File size: 83,485 Bytes
aefa98b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
[
 {
  "source": "NF-UNSW-NB15-v3",
  "mapping": {
   "source": "NF-UNSW-NB15-v3",
   "papers_read": [
    "2021_Sarhan_NetFlow_Datasets_for_Machine_Learning_Based_Network_Intrusio.pdf",
    "2021_Sarhan_Towards_a_Standard_Feature_Set_for_Network_Intrusion_Detecti.pdf",
    "2026_Shaik_LightGBM_Based_Intrusion_Detection_on_the_UNSW_NB15_Dataset_.pdf (searched only, no category descriptions)"
   ],
   "classes": [
    {
     "label": "Benign",
     "documented_behaviour": "Normal unmalicious flows; UNSW-NB15 benign traffic was generated by IXIA PerfectStorm as a hybrid of real benign activity and synthetic attacks, captured with tcpdump.",
     "evidence": "2021_Sarhan_NetFlow_Datasets...pdf p.8 Table 2: 'Benign ... Normal unmalicious flows'.",
     "techniques": [],
     "primary_tactic": "none (no ATT&CK mapping)",
     "confidence": "high",
     "netflow_observable": "yes",
     "caveat": "Not adversary behaviour; no ATT&CK technique applies. Neither paper describes NF-UNSW-NB15-v3; only v1 and v2 tables were read."
    },
    {
     "label": "Analysis",
     "documented_behaviour": "Group of threats against web applications via ports, emails and scripts. The papers name no tool or protocol.",
     "evidence": "'A group that presents a variety of threats that target web applications through ports, emails and scripts.' (2021_Sarhan_NetFlow_Datasets...pdf p.8, Table 2; same text in 2021_Sarhan_Towards...pdf p.6, Table 3)",
     "techniques": [
      {
       "id": "T1190",
       "name": "Exploit Public-Facing Application",
       "tactics": "initial-access",
       "fit": "primary"
      },
      {
       "id": "T1595.002",
       "name": "Vulnerability Scanning",
       "tactics": "reconnaissance",
       "fit": "secondary"
      }
     ],
     "primary_tactic": "initial-access",
     "confidence": "low",
     "netflow_observable": "partial",
     "caveat": "The description is a vague umbrella (web, email, ports, scripts). Several different behaviours are bundled, so one mapping cannot cover the class. Web payloads and email content are not visible in flow records. The papers do not show what the NF flows contain."
    },
    {
     "label": "Backdoor",
     "documented_behaviour": "Bypassing security mechanisms by replying to specially constructed client applications, which implies a hidden remote-access channel. No tool or protocol is named.",
     "evidence": "'A technique that aims to bypass security mechanisms by replying to specific constructed client applications.' (2021_Sarhan_NetFlow_Datasets...pdf p.8, Table 2)",
     "techniques": [
      {
       "id": "T1219",
       "name": "Remote Access Tools",
       "tactics": "command-and-control",
       "fit": "primary"
      },
      {
       "id": "T1571",
       "name": "Non-Standard Port",
       "tactics": "command-and-control",
       "fit": "secondary"
      }
     ],
     "primary_tactic": "command-and-control",
     "confidence": "low",
     "netflow_observable": "partial",
     "caveat": "The paper gives no mechanism. T1219 is inferred from the label and the phrase 'constructed client applications'. T1571 is a guess. Flows would show an unusual server port and a long-lived connection, but not the backdoor itself."
    },
    {
     "label": "DoS",
     "documented_behaviour": "Overloading a computer system's resources to prevent access to or availability of its data. No tool or protocol is named, so the layer (network or endpoint) is unknown.",
     "evidence": "'Denial of Service is an attempt to overload a computer system's resources with the aim of preventing access to or availability of its data.' (2021_Sarhan_NetFlow_Datasets...pdf p.8, Table 2)",
     "techniques": [
      {
       "id": "T1499",
       "name": "Endpoint Denial of Service",
       "tactics": "impact",
       "fit": "primary"
      },
      {
       "id": "T1498",
       "name": "Network Denial of Service",
       "tactics": "impact",
       "fit": "secondary"
      }
     ],
     "primary_tactic": "impact",
     "confidence": "medium",
     "netflow_observable": "yes",
     "caveat": "The class is certainly a DoS, but no sub-technique is supported because the flood type is not stated. Parent techniques are used. UNSW DoS is a different workload from BoT-IoT and CSE-CIC DoS. Mapping is per (source, label)."
    },
    {
     "label": "Exploits",
     "documented_behaviour": "Sequences of commands that control a host's behaviour through a known vulnerability. The specific vulnerabilities and services are not listed in the papers read.",
     "evidence": "'Are sequences of commands controlling the behaviour of a host through a known vulnerability.' (2021_Sarhan_NetFlow_Datasets...pdf p.8, Table 2)",
     "techniques": [
      {
       "id": "T1190",
       "name": "Exploit Public-Facing Application",
       "tactics": "initial-access",
       "fit": "primary"
      },
      {
       "id": "T1210",
       "name": "Exploitation of Remote Services",
       "tactics": "lateral-movement",
       "fit": "secondary"
      },
      {
       "id": "T1203",
       "name": "Exploitation for Client Execution",
       "tactics": "execution",
       "fit": "secondary"
      }
     ],
     "primary_tactic": "initial-access",
     "confidence": "medium",
     "netflow_observable": "partial",
     "caveat": "The papers do not say whether the targets are public-facing, internal or client-side, so T1190 versus T1210 is a judgment call. The class is exploit-based, but flows show only a connection to a service, not the exploit payload."
    },
    {
     "label": "Fuzzers",
     "documented_behaviour": "Sending large amounts of random data to crash a system and to discover security vulnerabilities. No tool is named.",
     "evidence": "'An attack in which the attacker sends large amounts of random data which cause a system to crash and also aim to discover security vulnerabilities in a system.' (2021_Sarhan_NetFlow_Datasets...pdf p.8, Table 2)",
     "techniques": [
      {
       "id": "T1595.002",
       "name": "Vulnerability Scanning",
       "tactics": "reconnaissance",
       "fit": "primary"
      },
      {
       "id": "T1499.004",
       "name": "Application or System Exploitation",
       "tactics": "impact",
       "fit": "secondary"
      }
     ],
     "primary_tactic": "reconnaissance",
     "confidence": "low",
     "netflow_observable": "partial",
     "caveat": "ATT&CK has no fuzzing technique. The description gives two goals, vulnerability discovery (reconnaissance) and crashing the target (impact), so the primary tactic is debatable. Random payloads are invisible in flows. Only many short flows or odd sizes could show."
    },
    {
     "label": "Generic",
     "documented_behaviour": "A method that targets cryptography and causes a collision with each block cipher. The papers give no tool, protocol or target.",
     "evidence": "'A method that targets cryptography and causes a collision with each block-cipher.' (2021_Sarhan_NetFlow_Datasets...pdf p.8, Table 2; 2021_Sarhan_Towards...pdf p.6, Table 3)",
     "techniques": [
      {
       "id": "T1600.001",
       "name": "Reduce Key Space",
       "tactics": "defense-impairment",
       "fit": "primary"
      },
      {
       "id": "T1600",
       "name": "Weaken Encryption",
       "tactics": "defense-impairment",
       "fit": "secondary"
      }
     ],
     "primary_tactic": "defense-impairment",
     "confidence": "low",
     "netflow_observable": "no",
     "caveat": "There is no good ATT&CK counterpart. Cryptanalytic collision attacks are not an enterprise adversary technique, and T1600 describes weakening a target's own crypto on a network device. This is a forced approximation for completeness only. The behaviour is not observable in flow statistics, so the class is likely distinguished by protocol and port artefacts of the testbed rather than by attack semantics."
    },
    {
     "label": "Reconnaissance",
     "documented_behaviour": "Gathering information about a network host, also known as a probe. No tool is named in the papers read.",
     "evidence": "'A technique for gathering information about a network host and is also known as a probe.' (2021_Sarhan_NetFlow_Datasets...pdf p.8, Table 2)",
     "techniques": [
      {
       "id": "T1595",
       "name": "Active Scanning",
       "tactics": "reconnaissance",
       "fit": "primary"
      },
      {
       "id": "T1046",
       "name": "Network Service Discovery",
       "tactics": "discovery",
       "fit": "secondary"
      }
     ],
     "primary_tactic": "reconnaissance",
     "confidence": "medium",
     "netflow_observable": "yes",
     "caveat": "The papers do not say whether the probing is port scanning, host sweeping or vulnerability scanning, so the parent T1595 is used. T1595.001 and T1595.002 are not supported. Probing shows up in flows as many short flows to many ports or hosts."
    },
    {
     "label": "Shellcode",
     "documented_behaviour": "Malware that penetrates code to control a victim's host. This describes a payload, with no delivery mechanism, tool or protocol given.",
     "evidence": "'A malware that penetrates a code to control a victim's host.' (2021_Sarhan_NetFlow_Datasets...pdf p.8, Table 2)",
     "techniques": [
      {
       "id": "T1203",
       "name": "Exploitation for Client Execution",
       "tactics": "execution",
       "fit": "primary"
      },
      {
       "id": "T1059",
       "name": "Command and Scripting Interpreter",
       "tactics": "execution",
       "fit": "secondary"
      },
      {
       "id": "T1055",
       "name": "Process Injection",
       "tactics": "stealth, privilege-escalation",
       "fit": "secondary"
      }
     ],
     "primary_tactic": "execution",
     "confidence": "low",
     "netflow_observable": "partial",
     "caveat": "Shellcode is a payload, not a delivery step, and the description is a loose paraphrase. The mapping is inferred from the label. The delivery vector is probably an exploit, so the class overlaps with Exploits. Payload bytes are invisible in flows, though the flow count is very small (1,365 in v1)."
    },
    {
     "label": "Worms",
     "documented_behaviour": "Attacks that replicate themselves and spread to other computers. No worm family or propagation protocol is named.",
     "evidence": "'Attacks that replicate themselves and spread to other computers.' (2021_Sarhan_NetFlow_Datasets...pdf p.8, Table 2)",
     "techniques": [
      {
       "id": "T1210",
       "name": "Exploitation of Remote Services",
       "tactics": "lateral-movement",
       "fit": "primary"
      },
      {
       "id": "T1021",
       "name": "Remote Services",
       "tactics": "lateral-movement",
       "fit": "secondary"
      }
     ],
     "primary_tactic": "lateral-movement",
     "confidence": "medium",
     "netflow_observable": "partial",
     "caveat": "Self-propagation is documented, but the propagation method is not, so T1210 is the usual worm mechanism rather than something the paper states. Spread appears in flows as one host contacting many hosts, but it is rare (153 flows in v1) and hard to separate from Exploits. ATT&CK has no dedicated worm technique."
    }
   ]
  },
  "challenge": {
   "source": "NF-UNSW-NB15-v3",
   "verdicts": [
    {
     "label": "Benign",
     "mapping_ok": true,
     "problem": "No refutation. The quote 'Normal unmalicious flows' is verbatim in 2021_Sarhan_NetFlow_Datasets p.8 Table 2. The IXIA PerfectStorm, tcpdump and hybrid real/synthetic traffic statement is also in that paper (introduction, p.3-4, not cited by page). No technique is correct. The caveat that v3 is not described is accurate.",
     "confidence_after": "high"
    },
    {
     "label": "Analysis",
     "mapping_ok": false,
     "problem": "The quote is verbatim (NetFlow_Datasets p.8 Table 2; Towards p.6 Table 3). Secondary T1595.002 is not supported. ATT&CK describes it as checking software and version against a specific exploit, and the papers mention no vulnerability scanning. The row is also inconsistent with the Reconnaissance row, which rejects T1595.00x as unsupported. The text only says 'ports', so the parent T1595 is the most that can be defended. T1190 as primary fits 'target web applications' loosely and stays low. Observability 'partial' is fine. The original UNSW-NB15 paper (port scan, spam, html penetration) is not in the corpus, so do not cite it.",
     "corrected_techniques": "T1190 Exploit Public-Facing Application (initial-access, primary); T1595 Active Scanning (reconnaissance, secondary)",
     "corrected_primary_tactic": "initial-access",
     "confidence_after": "low"
    },
    {
     "label": "Backdoor",
     "mapping_ok": false,
     "problem": "The quote is verbatim (NetFlow_Datasets p.8 Table 2). T1219 Remote Access Tools is the wrong technique. ATT&CK scopes it to legitimate remote-access software such as AnyDesk, and the paper describes a custom backdoor that replies only to specially constructed clients. The behaviour matches T1205 Traffic Signaling more closely: a magic value or special request triggers a response from hidden functionality, and the STIX tactics are stealth, persistence and command-and-control. T1571 is admitted as a guess with no support in the text, so drop it. Keep the primary tactic as command-and-control. Observability 'partial' is fine.",
     "corrected_techniques": "T1205 Traffic Signaling (stealth, persistence, command-and-control; primary)",
     "corrected_primary_tactic": "command-and-control",
     "confidence_after": "low"
    },
    {
     "label": "DoS",
     "mapping_ok": true,
     "problem": "No refutation. The quote is verbatim (NetFlow_Datasets p.8 Table 2). 'Overload a computer system's resources' fits T1499 (system-resource exhaustion) better than T1498 (bandwidth exhaustion). Both are valid, non-deprecated, impact-tactic parent techniques, and the sub-techniques (T1499.001-.004, T1498.001-.002) are correctly left out because the flood type is not stated. Observability 'yes' is fine. Medium confidence is acceptable for the tactic, but the technique-level choice between endpoint and network is inferred.",
     "confidence_after": "medium"
    },
    {
     "label": "Exploits",
     "mapping_ok": false,
     "problem": "The quote is verbatim (NetFlow_Datasets p.8 Table 2). T1203 Exploitation for Client Execution is unsupported. The paper says 'a host', not a client application, and ATT&CK scopes T1203 to client software. The primary tactic is also not determined by the papers. T1190 needs an Internet-facing target, T1210 an internal one, and T1068 would be privilege escalation. The row's own caveat calls this a judgment call, so 'medium' is too high. T1190 stays as the conventional choice for server-side exploitation of a known vulnerability, but confidence is low. Observability 'partial' is fine.",
     "corrected_techniques": "T1190 Exploit Public-Facing Application (initial-access, primary); T1210 Exploitation of Remote Services (lateral-movement, secondary)",
     "corrected_primary_tactic": "initial-access",
     "confidence_after": "low"
    },
    {
     "label": "Fuzzers",
     "mapping_ok": false,
     "problem": "The quote is verbatim (NetFlow_Datasets p.8 Table 2), but the primary technique is questionable. The sentence leads with 'random data which cause a system to crash' and adds 'also aim to discover security vulnerabilities', so the crash is the stated effect. T1499.004 (exploit vulnerabilities to crash an application or system, impact) matches that clause directly. ATT&CK describes T1595.002 as scanning for known vulnerabilities by software and version against a specific exploit. That does not match fuzzing, which looks for unknown flaws. Swap the order: T1499.004 primary and T1595.002 secondary, and note that ATT&CK has no fuzzing technique. The primary tactic is genuinely ambiguous, so confidence stays low. Observability 'partial' is fine.",
     "corrected_techniques": "T1499.004 Application or System Exploitation (impact, primary); T1595.002 Vulnerability Scanning (reconnaissance, secondary)",
     "corrected_primary_tactic": "impact",
     "confidence_after": "low"
    },
    {
     "label": "Generic",
     "mapping_ok": false,
     "problem": "The quote is verbatim (NetFlow_Datasets p.8 Table 2; Towards p.6 Table 3). The mapping is a forced fit. T1600.001 Reduce Key Space covers weakening the cipher strength of a compromised network device, which is not a cryptanalytic collision attack on block ciphers. Listing the sub-technique as primary and its own parent as secondary is redundant. Mapping 5,570 flows to defense-impairment would wrongly place this class in a family with real adversary behaviour. The honest result is no ATT&CK counterpart, as for Benign, with the class kept as an unmapped label. Observability 'no' is fine.",
     "corrected_techniques": "none (no ATT&CK technique; keep as an unmapped class, evidence 'cryptographic collision attack, not an enterprise adversary technique')",
     "corrected_primary_tactic": "none (no ATT&CK mapping)",
     "confidence_after": "low"
    },
    {
     "label": "Reconnaissance",
     "mapping_ok": true,
     "problem": "No refutation. The quote is verbatim (NetFlow_Datasets p.8 Table 2). The 'probe' wording supports T1595 Active Scanning (reconnaissance) at parent level, and the row rightly rejects the sub-techniques. T1046 as secondary is acceptable but marginal, since in ATT&CK it is a discovery step usually run from inside a network. Observability 'yes' and medium confidence are fine.",
     "confidence_after": "medium"
    },
    {
     "label": "Shellcode",
     "mapping_ok": false,
     "problem": "The quote is verbatim (NetFlow_Datasets p.8 Table 2). T1203 Exploitation for Client Execution is unsupported as primary, because the paper says nothing about a client application. T1203 would also overlap wrongly with the Exploits row. ATT&CK has no shellcode technique. 'Control a victim's host' only supports the execution tactic, with T1059 as the closest generic technique. T1055 is a weak secondary reading of 'penetrates a code'. The delivery step (T1190 or T1210) belongs to the Exploits row. Confidence stays low. Observability 'partial' is arguably 'no', since only payload bytes distinguish the class, but 'partial' is defensible given the small flow count.",
     "corrected_techniques": "T1059 Command and Scripting Interpreter (execution, primary); T1055 Process Injection (stealth, privilege-escalation; secondary)",
     "corrected_primary_tactic": "execution",
     "confidence_after": "low"
    },
    {
     "label": "Worms",
     "mapping_ok": false,
     "problem": "The quote is verbatim (NetFlow_Datasets p.8 Table 2). 'Spread to other computers' does support lateral-movement as the tactic. T1021 Remote Services is unsupported as secondary, because ATT&CK describes it as logging in with valid accounts, and a worm needs no credentials. T1570 Lateral Tool Transfer fits 'replicate themselves' (copying files between systems) and is non-deprecated in STIX. T1210 as primary is a standard but unstated inference. Confidence 'medium' is acceptable for the tactic, but the technique is low. Observability 'partial' is fine.",
     "corrected_techniques": "T1210 Exploitation of Remote Services (lateral-movement, primary); T1570 Lateral Tool Transfer (lateral-movement, secondary)",
     "corrected_primary_tactic": "lateral-movement",
     "confidence_after": "medium"
    }
   ]
  }
 },
 {
  "source": "NF-ToN-IoT-v3",
  "mapping": {
   "source": "NF-ToN-IoT-v3",
   "papers_read": [
    "2020_Alsaedi_TON_IoT_Telemetry_Dataset_A_New_Generation_Dataset_of_IoT_an.pdf (attack descriptions, journal pp. 165137-165140)",
    "2021_Sarhan_NetFlow_Datasets_for_Machine_Learning_Based_Network_Intrusio.pdf (Table 4 class definitions, p. 9; NF-ToN-IoT generation)",
    "2021_Sarhan_Towards_a_Standard_Feature_Set_for_Network_Intrusion_Detecti.pdf (ToN-IoT/NF-v2 description only; no per-class text)",
    "2026_Uddin_BigFlow_NIDS_dataset_DataInBrief.pdf (grepped for ToN-IoT class text; none found)"
   ],
   "classes": [
    {
     "label": "Backdoor",
     "documented_behaviour": "Metasploitable3 framework used from two Kali attacker VMs (192.168.1.33, .37) to gain unauthorised remote access to infected IIoT devices via backdoor malware; devices could then be controlled and enrolled in botnets.",
     "evidence": "\"allows an adversary to gain unauthorised remote access to the infected IIoT devices by a backdoor malware... We utilised the Metasploitable3 framework\" (Alsaedi 2020, p.165139)",
     "techniques": [
      {
       "id": "T1219",
       "name": "Remote Access Tools",
       "tactics": "command-and-control",
       "fit": "primary"
      },
      {
       "id": "T1190",
       "name": "Exploit Public-Facing Application",
       "tactics": "initial-access",
       "fit": "secondary"
      }
     ],
     "primary_tactic": "command-and-control",
     "confidence": "medium",
     "netflow_observable": "partial",
     "caveat": "ATT&CK has no generic 'backdoor' technique; the paper names no specific backdoor payload or protocol, so T1219 is an approximation and T1190 is inferred from 'Metasploitable3 exploit'. Flows show only connections to the backdoor port, not the control activity. Sarhan Table 4 gives a different generic definition (p.9)."
    },
    {
     "label": "DDoS",
     "documented_behaviour": "Python scripts (ddos.py, ddos2.py, ddos2_broker.py) driving the UFONet toolkit from Kali VMs 192.168.1.30, .31 and .34-.38, flooding IIoT devices/MQTT broker with many connections to exhaust CPU and memory.",
     "evidence": "\"different python scripts (e.g, ddos.py, ddos2.py and ddos2_broker.py) were written to launch DDoS attacks using the UFONet toolkit\" (Alsaedi 2020, p.165139)",
     "techniques": [
      {
       "id": "T1499.002",
       "name": "Service Exhaustion Flood",
       "tactics": "impact",
       "fit": "primary"
      },
      {
       "id": "T1498.001",
       "name": "Direct Network Flood",
       "tactics": "impact",
       "fit": "secondary"
      }
     ],
     "primary_tactic": "impact",
     "confidence": "medium",
     "netflow_observable": "yes",
     "caveat": "The papers name UFONet but do not say which protocol layer it floods; preferring the service-level sub-technique rests on my background knowledge of UFONet (HTTP floods), not on the papers. The 'distributed' sources are ~8 Kali VMs, not a real botnet. DDoS is the largest class (326,345 flows, Sarhan p.9) and shares IPs with DoS."
    },
    {
     "label": "DoS",
     "documented_behaviour": "A Scapy-based Python script (dos.py) run from Kali VMs 192.168.1.30, .31 and .39 sending floods of connection attempts against IoT/IIoT devices to disrupt service.",
     "evidence": "\"a python script (e.g., dos.py) was developed to perform DoS attacks using the Scapy package\" (Alsaedi 2020, p.165139)",
     "techniques": [
      {
       "id": "T1499",
       "name": "Endpoint Denial of Service",
       "tactics": "impact",
       "fit": "primary"
      },
      {
       "id": "T1498.001",
       "name": "Direct Network Flood",
       "tactics": "impact",
       "fit": "secondary"
      }
     ],
     "primary_tactic": "impact",
     "confidence": "medium",
     "netflow_observable": "yes",
     "caveat": "The paper does not state the packet type (SYN, UDP, ICMP) or the target layer, so no endpoint sub-technique is supported; the parent T1499 is used deliberately. Bandwidth versus resource exhaustion cannot be told apart from the text."
    },
    {
     "label": "injection",
     "documented_behaviour": "Two bash scripts (injection-1.sh, injection-2.sh) injecting malicious data/code inputs into web applications (DVWA, a vulnerable public PHP site, Security Shepherd) and IoT service web pages from Kali VMs 192.168.1.{30,31,33,35,36,38}. Sarhan names SQL and code injection as the main kinds.",
     "evidence": "\"two bash scripts (e.g., injection-1.sh and injection-2.sh) were written to inject data inputs against web applications (e.g., DVWA... and Security Shepherd VMs\" (Alsaedi 2020, p.165139)",
     "techniques": [
      {
       "id": "T1190",
       "name": "Exploit Public-Facing Application",
       "tactics": "initial-access",
       "fit": "primary"
      },
      {
       "id": "T1565.002",
       "name": "Transmitted Data Manipulation",
       "tactics": "impact",
       "fit": "secondary"
      }
     ],
     "primary_tactic": "initial-access",
     "confidence": "medium",
     "netflow_observable": "partial",
     "caveat": "The injection payloads are in HTTP bodies, which NetFlow cannot see; only the repeated connections to the web servers show. T1565.002 reflects the paper's claim that injection can manipulate telemetry and control commands, but no such manipulation is shown. Injection is the largest attack class after DDoS (468,539 flows) and is likely dominated by automated scripts."
    },
    {
     "label": "mitm",
     "documented_behaviour": "Ettercap run from Kali VMs 192.168.1.31-34 to carry out ARP cache poisoning, ICMP redirection and port-stealing, intercepting communication between devices and possibly manipulating it.",
     "evidence": "\"We used the Ettercap tool to execute ARP Cache poisoning, ICMP redirection and port-stealing attacks.\" (Alsaedi 2020, p.165140)",
     "techniques": [
      {
       "id": "T1557.002",
       "name": "ARP Cache Poisoning",
       "tactics": "credential-access, collection",
       "fit": "primary"
      },
      {
       "id": "T1557",
       "name": "Adversary-in-the-Middle",
       "tactics": "credential-access, collection",
       "fit": "secondary"
      },
      {
       "id": "T1040",
       "name": "Network Sniffing",
       "tactics": "credential-access, discovery",
       "fit": "secondary"
      }
     ],
     "primary_tactic": "credential-access",
     "confidence": "high",
     "netflow_observable": "no",
     "caveat": "ARP poisoning is layer 2 and is not exported as IP flows; the labelled flows are ordinary IP traffic in the attack window, so label noise is likely. ICMP redirect and port stealing are covered only by the parent T1557. T1040 is inferred from interception, not stated. Only 1,295 flows."
    },
    {
     "label": "password",
     "documented_behaviour": "Dictionary attacks using CeWL-generated wordlists (password-1.sh) and brute-force attacks using Hydra (password-2.sh), run in parallel against IoT/IIoT devices from Kali VMs 192.168.1.30, .31, .33, .35, .38.",
     "evidence": "\"password-1.sh using the CeWL toolkit for dictionary attacks and (password-2.sh) using the Hydra toolkit for brute force attacks\" (Alsaedi 2020, p.165140)",
     "techniques": [
      {
       "id": "T1110.001",
       "name": "Password Guessing",
       "tactics": "credential-access",
       "fit": "primary"
      }
     ],
     "primary_tactic": "credential-access",
     "confidence": "high",
     "netflow_observable": "partial",
     "caveat": "Online guessing shows as many short flows to one service port, but flows cannot reveal success or the target protocol's credentials. Offline cracking (T1110.002) is not documented. Sarhan Table 4 also says 'or sniffing' (p.9); the Alsaedi paper documents only guessing, so T1040 is not mapped."
    },
    {
     "label": "ransomware",
     "documented_behaviour": "Kali VMs 192.168.1.33 and .37 used Metasploitable3 to exploit weaknesses in target devices and then ran a ransomware attack that denies access by encrypting and demanding payment.",
     "evidence": "\"We used a Metasploitable3 framework to exploit weaknesses in the target devices and then executed the ransomware attack.\" (Alsaedi 2020, p.165139)",
     "techniques": [
      {
       "id": "T1486",
       "name": "Data Encrypted for Impact",
       "tactics": "impact",
       "fit": "primary"
      },
      {
       "id": "T1190",
       "name": "Exploit Public-Facing Application",
       "tactics": "initial-access",
       "fit": "secondary"
      }
     ],
     "primary_tactic": "impact",
     "confidence": "medium",
     "netflow_observable": "no",
     "caveat": "File encryption is host-level and leaves no flow signature; any flows labelled ransomware are the exploitation or delivery traffic. The paper names no ransomware family or exploit, so T1190 is inferred. Very rare (142 flows, Sarhan p.9), so flow-level evaluation will be weak. Shares attacker IPs with Backdoor."
    },
    {
     "label": "scanning",
     "documented_behaviour": "Nmap and Nessus run from Kali VMs 192.168.1.30-33 and .38 against the IoT/IIoT devices in 192.168.1.0/24 and the public MQTT broker to find open ports and services.",
     "evidence": "\"We used both Nmap and Nessus tools from the offensive Kali systems to perform scanning attacks against the victim IoT/IIoT devices in the subnet (192.168.1.0/24)\" (Alsaedi 2020, p.165137)",
     "techniques": [
      {
       "id": "T1046",
       "name": "Network Service Discovery",
       "tactics": "discovery",
       "fit": "primary"
      },
      {
       "id": "T1595.002",
       "name": "Vulnerability Scanning",
       "tactics": "reconnaissance",
       "fit": "secondary"
      },
      {
       "id": "T1595.001",
       "name": "Scanning IP Blocks",
       "tactics": "reconnaissance",
       "fit": "secondary"
      }
     ],
     "primary_tactic": "discovery",
     "confidence": "high",
     "netflow_observable": "yes",
     "caveat": "Tactic depends on vantage: attackers sit inside the testbed subnet, so T1046 (discovery) is primary; an external attacker would be T1595 (reconnaissance). Nessus supports T1595.002, but flows cannot separate Nmap from Nessus. The paper's scan against the MQTT broker is documented, with no further detail."
    },
    {
     "label": "xss",
     "documented_behaviour": "Scripts xss1.sh and xss2.sh using the XSSer toolkit to inject web scripts (HTTP/JavaScript) into DVWA, Security Shepherd and IoT service web pages from Kali VMs 192.168.1.{32,35,36,39}.",
     "evidence": "\"scripts (e.g., xss1-sh and xss2.sh) using the Cross-Site Scripter (XSSer) toolkit to perform XSS attacks on the Web applications of DVWA, Security Shepherd VMs\" (Alsaedi 2020, p.165140)",
     "techniques": [
      {
       "id": "T1190",
       "name": "Exploit Public-Facing Application",
       "tactics": "initial-access",
       "fit": "primary"
      },
      {
       "id": "T1059.007",
       "name": "JavaScript",
       "tactics": "execution",
       "fit": "secondary"
      }
     ],
     "primary_tactic": "initial-access",
     "confidence": "medium",
     "netflow_observable": "partial",
     "caveat": "ATT&CK has no XSS technique. The documented behaviour is the tool firing payloads at a vulnerable web server, so T1190 fits; T1189 (drive-by) was rejected because no victim browsing is described. Payloads are invisible to NetFlow, so xss and injection flows will look alike (both are web attacks on the same VMs). 99,944 flows."
    },
    {
     "label": "Benign",
     "documented_behaviour": "Normal unmalicious flows from the IoT/IIoT testbed (MQTT telemetry, simulated fridge, garage door, GPS, motion light, Modbus, thermostat, weather).",
     "evidence": "\"Normal unmalicious flows\" (Sarhan 2021 NetFlow Datasets, Table 4, p.9)",
     "techniques": [],
     "primary_tactic": "none (not an adversary behaviour)",
     "confidence": "high",
     "netflow_observable": "yes",
     "caveat": "Benign is not an ATT&CK concept and has no mapping. Testbed benign traffic is synthetic and from different hosts than the attackers, which matters for the source and IP-shortcut concerns raised in earlier steps."
    }
   ]
  },
  "challenge": {
   "source": "NF-ToN-IoT-v3",
   "verdicts": [
    {
     "label": "Backdoor",
     "mapping_ok": false,
     "problem": "All IDs exist and are not deprecated (STIX v20), but T1219 is wrong. Its STIX description says adversaries use legitimate remote access tools (TeamViewer-style), and the paper describes backdoor malware, not legitimate software. The paper only documents Metasploitable3 exploitation from Kali .33/.37, so the one step it supports is T1190. Controlling the device afterwards is stated only generically. The confidence is too high for that, so use low. Sarhan Table 4 (p.9) defines Backdoor differently ('replying to specific constructed client applications'), so the sources disagree. T1190 assumes the targets are public-facing, which is not shown for devices inside the subnet. T1210 would be the in-LAN alternative and the paper cannot decide between them. Quote and page (Alsaedi p.165139) are accurate; 'partial' observability is fine. The BigFlow Data-in-Brief Table 2 (merged label, 'Unauthorized remote access via hidden backdoor') was not noted, although the papers_read entry says it had no class text.",
     "confidence_after": "low",
     "corrected_primary_tactic": "initial-access",
     "corrected_techniques": "T1190 Exploit Public-Facing Application (primary, initial-access); drop T1219. Optionally note T1210 Exploitation of Remote Services (lateral-movement) as an in-subnet alternative. State that ATT&CK has no backdoor technique and the C2 behaviour is not documented."
    },
    {
     "label": "DDoS",
     "mapping_ok": false,
     "problem": "The primary T1499.002 is not supported by the papers. Alsaedi p.165139 says only 'flooding ... with numerous connections so to exhaust the device resources (CPU and memory)'. That matches parent T1499 Endpoint Denial of Service, whose STIX text covers exhausting system resources without saturating the network. T1499.002 (STIX: HTTP/DNS/web service floods) rests on outside knowledge of UFONet, which the rules forbid. Make T1499 primary and list T1499.002 and T1498.001 as low-confidence secondaries. Two further fixes. First, the caveat 'DDoS is the largest class (326,345)' is wrong. Sarhan Table 4 gives Injection 468,539 > DDoS 326,345 > Password 156,299. Second, the DDoS row is internally inconsistent with the DoS row, which uses the parent T1499 for the same lack of layer evidence. IDs exist and are not deprecated. Quote and p.165139 are accurate and the IP lists match the text. BigFlow Table 2 describes DDoS only as 'exhaust network or server resources', so network versus endpoint is undecided. Tactic impact and observability 'yes' are fine.",
     "confidence_after": "medium",
     "corrected_primary_tactic": "impact",
     "corrected_techniques": "T1499 Endpoint Denial of Service (primary, impact); T1499.002 Service Exhaustion Flood (secondary, low, tool-knowledge only); T1498.001 Direct Network Flood (secondary, low)"
    },
    {
     "label": "DoS",
     "mapping_ok": true,
     "problem": "Survives refutation. T1499 and T1498.001 exist, are not deprecated and have tactic impact. The quote and p.165139 are accurate, and the IPs .30/.31/.39 match. Using the parent T1499 is correct because the paper gives no packet type or layer. Minor: T1498.001 (high-volume traffic; UDP/ICMP/TCP) is not supported by the paper, which says nothing about bandwidth, so keep it as a low-confidence secondary. It is not a primary mapping. Sarhan Table 4's 'overload a computer system's resources' also fits T1499. Observability 'yes' and medium confidence are fine.",
     "confidence_after": "medium"
    },
    {
     "label": "injection",
     "mapping_ok": false,
     "problem": "T1190 as primary is well supported: untrusted input to web apps (DVWA, Security Shepherd), SQL/code injection per Sarhan p.9, and the quote at p.165139 is accurate. The secondary T1565.002 is wrong. Per STIX it means altering data en route (requires interception or AiTM), whereas injection into a web form is not transit manipulation. The paper's claim that injection 'can manipulate telemetry data and control commands' is generic and unshown, so at most the parent T1565 Data Manipulation fits, at low confidence. The caveat 'largest attack class after DDoS' is also wrong. Injection (468,539) is the largest class and DDoS (326,345) is second (Sarhan Table 4, p.9). Those counts are for v1 NF-ToN-IoT, so check them against the v3 data. Observability 'partial' and medium confidence are fine.",
     "confidence_after": "medium",
     "corrected_primary_tactic": "initial-access",
     "corrected_techniques": "T1190 Exploit Public-Facing Application (primary, initial-access); T1565 Data Manipulation (secondary, impact, low); drop T1565.002"
    },
    {
     "label": "mitm",
     "mapping_ok": false,
     "problem": "IDs are valid and not deprecated. The quote and p.165140 are accurate (Ettercap, ARP cache poisoning, ICMP redirect, port stealing; IPs .31-.34). T1557.002 is directly named, so high confidence for the technique is defensible. One judgement is wrong: NetFlow observability 'no' should be 'partial'. ICMP redirect is ordinary ICMP/IP traffic that NetFlow records (ICMP type/code), so only ARP poisoning and port stealing (L2) are invisible. Also note the label aggregates three methods and the paper does not say which one produced the labelled flows, so T1557 parent is an equally good primary at the label level. T1040 as a secondary is an inference, which the row already says. Primary tactic credential-access is valid (T1557 lists credential-access and collection).",
     "confidence_after": "high",
     "corrected_primary_tactic": "credential-access",
     "corrected_techniques": "Keep T1557.002 ARP Cache Poisoning, T1557 Adversary-in-the-Middle, T1040 Network Sniffing; set netflow_observable to 'partial' (ICMP redirect visible; ARP and port stealing not)."
    },
    {
     "label": "password",
     "mapping_ok": true,
     "problem": "Survives refutation. T1110.001 exists, is not deprecated and is credential-access. Hydra brute force and CeWL dictionary attacks to guess device passwords fit online password guessing. The quote omits '(e.g.,' and citation markers but is faithful, and p.165140 is correct (IPs .30/.31/.33/.35/.38 match). T1110.002 and T1110.003 are rightly not mapped, because the paper never describes offline cracking or spraying. Sarhan's 'or sniffing' (p.9) is correctly not mapped. The paper heading says 'Password Cracking', but its text says 'to guess the password'. Optional: add the parent T1110 Brute Force for the aggregate label. Observability 'partial' and high confidence are acceptable.",
     "confidence_after": "high"
    },
    {
     "label": "ransomware",
     "mapping_ok": true,
     "problem": "Survives refutation. T1486 exists, is not deprecated and is impact. The quote is exact and the page is correct (Alsaedi p.165139, Kali .33/.37). Paper and Sarhan p.9 both state the encrypt-and-demand-payment behaviour. T1190 is an inference, because the paper only says 'exploit weaknesses in the target devices'. It is still reasonable, but for in-subnet targets T1210 is the alternative, and the row should say T1190 is low confidence. Observability 'no' and the 142-flow count (Sarhan p.9) are right. Medium confidence is appropriate, since encryption is stated generically and not shown.",
     "confidence_after": "medium"
    },
    {
     "label": "scanning",
     "mapping_ok": false,
     "problem": "IDs are valid and not deprecated, and the quote and p.165137 are accurate. The primary tactic is wrong. The paper itself frames scanning as 'the first step where an attacker gathers information about a target system ... before launching the actual attack' (p.165137), which is reconnaissance. The STIX T1046 text describes discovery from tools 'brought onto a system', i.e. post-compromise. Here the Kali VMs are the attacker's own machines probing victim infrastructure, which is exactly the T1595 definition ('adversary probes victim infrastructure via network traffic'). The in-subnet placement is a lab artefact, and 'discovery' would also split this class from other sources' scanning/reconnaissance labels. Make reconnaissance primary with T1595 (T1595.001 for the /24 sweep, T1595.002 for Nessus) and keep T1046 as secondary for port/service discovery. Confidence 'high' is too high for the sub-technique split, because flows cannot separate Nmap from Nessus. Use medium for the sub-techniques and keep the tactic call medium-high. Observability 'yes' is fine.",
     "confidence_after": "medium",
     "corrected_primary_tactic": "reconnaissance",
     "corrected_techniques": "T1595 Active Scanning (primary, reconnaissance); T1595.001 Scanning IP Blocks; T1595.002 Vulnerability Scanning; T1046 Network Service Discovery (secondary, discovery)"
    },
    {
     "label": "xss",
     "mapping_ok": true,
     "problem": "Survives refutation. T1190 and T1059.007 exist, are not deprecated and have tactics initial-access and execution. The quote is faithful and the page is correct (XSSer against DVWA/Security Shepherd, IPs .32/.35/.36/.39, p.165140; 99,944 flows per Sarhan p.9). The paper says XSS 'attempts to run malicious commands on a Web server', which supports T1190. T1189 is correctly rejected because no victim browsing is shown, although Sarhan's 'send malicious scripts to end-users' hints at client-side impact. T1059.007 is only a weak secondary because the script runs in the victim browser and nothing in the papers shows execution. Medium confidence and 'partial' observability are fine.",
     "confidence_after": "medium"
    },
    {
     "label": "Benign",
     "mapping_ok": true,
     "problem": "Survives refutation. Benign is not an adversary behaviour, so an empty technique list is correct. The quote 'Normal unmalicious flows' is accurate (Sarhan Table 4, p.9). The device list in documented_behaviour comes from Alsaedi's testbed description and is not cited for Benign, which is only a minor provenance point.",
     "confidence_after": "high"
    }
   ]
  }
 },
 {
  "source": "NF-BoT-IoT-v3",
  "mapping": {
   "source": "NF-BoT-IoT-v3",
   "papers_read": [
    "2019_Koroniotis_Towards_the_development_of_realistic_botnet_dataset_in_the_I.pdf (Section 4.1 and Table 5, PDF pp.18-20)",
    "2021_Sarhan_NetFlow_Datasets_for_Machine_Learning_Based_Network_Intrusio.pdf (NF-BoT-IoT Table 3, PDF p.8)",
    "2021_Sarhan_Towards_a_Standard_Feature_Set_for_Network_Intrusion_Detecti.pdf (extracted, not used as evidence)"
   ],
   "classes": [
    {
     "label": "DDoS",
     "documented_behaviour": "Bot-IoT testbed: DDoS from bots against the Ubuntu server using hping3 (TCP SYN flood, UDP flood) and GoldenEye (HTTP POST flood). Table 5: TCP 19.5M and UDP 19.0M records vs HTTP 19.8K, so about 99.9% is volumetric TCP/UDP flooding.",
     "evidence": "Koroniotis 2019 PDF p.19: \"we used the tool Hping3 for both DDoS, DoS for TCP and UDP, for example hping3 --syn --flood -d 100 -p 80 192.168.100.3\"; HTTP via Golden-eye (Table 5, p.20).",
     "techniques": [
      {
       "id": "T1498.001",
       "name": "Direct Network Flood",
       "tactics": "impact",
       "fit": "primary"
      },
      {
       "id": "T1499.001",
       "name": "OS Exhaustion Flood",
       "tactics": "impact",
       "fit": "secondary"
      },
      {
       "id": "T1499.003",
       "name": "Application Exhaustion Flood",
       "tactics": "impact",
       "fit": "secondary"
      }
     ],
     "primary_tactic": "impact",
     "confidence": "high",
     "netflow_observable": "yes",
     "caveat": "ATT&CK has no DoS vs DDoS distinction, so both labels map to the same techniques; the difference is only the number of sources (a dataset property, worth checking against source-IP counts). A SYN flood is arguably T1499.001 rather than T1498.001, a judgement call. The HTTP part is tiny. Tool-generated lab floods, not realistic traffic."
    },
    {
     "label": "DoS",
     "documented_behaviour": "Same tools and protocols as DDoS (hping3 TCP SYN and UDP floods; GoldenEye HTTP) but from a single attacker. Table 5: TCP 12.3M, UDP 20.7M, HTTP 29.7K records.",
     "evidence": "Koroniotis 2019 PDF p.19: \"Denial of Service ... attempt to disrupt a service, thus making it unavailable to legitimate users\"; tools hping3 and Golden-eye (Table 5, p.20).",
     "techniques": [
      {
       "id": "T1498.001",
       "name": "Direct Network Flood",
       "tactics": "impact",
       "fit": "primary"
      },
      {
       "id": "T1499.001",
       "name": "OS Exhaustion Flood",
       "tactics": "impact",
       "fit": "secondary"
      },
      {
       "id": "T1499.003",
       "name": "Application Exhaustion Flood",
       "tactics": "impact",
       "fit": "secondary"
      }
     ],
     "primary_tactic": "impact",
     "confidence": "high",
     "netflow_observable": "yes",
     "caveat": "Behaviourally identical to DDoS in ATT&CK terms; separating them in a classifier relies on source multiplicity, not on a different technique. The paper text says DDoS and DoS are both 'performed by a group of compromised machines called Bots', which blurs the distinction, so treat the DoS/DDoS boundary as unreliable."
    },
    {
     "label": "Reconnaissance",
     "documented_behaviour": "Active probing: port/service scanning with Nmap (-sT) and hping3 (SYN scan), about 1.46M records, and OS fingerprinting with Nmap -O and Xprobe2, about 0.36M records. Passive probing was excluded.",
     "evidence": "Koroniotis 2019 PDF p.18: \"Nmap was launched to scan open services and protocols of the VMs, for example, nmap -sT 192.168.100.3\"; OS fingerprinting with Nmap and Xprobe2.",
     "techniques": [
      {
       "id": "T1595",
       "name": "Active Scanning",
       "tactics": "reconnaissance",
       "fit": "primary"
      },
      {
       "id": "T1046",
       "name": "Network Service Discovery",
       "tactics": "discovery",
       "fit": "secondary"
      },
      {
       "id": "T1592.002",
       "name": "Software",
       "tactics": "reconnaissance",
       "fit": "secondary"
      }
     ],
     "primary_tactic": "reconnaissance",
     "confidence": "high",
     "netflow_observable": "yes",
     "caveat": "The label mixes service scanning (about 80%) and OS fingerprinting (about 20%); NF-BoT-IoT does not separate them. T1595.001 (Scanning IP Blocks) was not chosen because single hosts were scanned, not ranges. T1046 is tagged Discovery (post-compromise) while the scans here run from outside, so T1595 is the better fit. T1592.002 covers OS only loosely."
    },
    {
     "label": "Theft",
     "documented_behaviour": "Data theft: Metasploit exploits EternalBlue (SMB) on Windows 7 and weak admin credentials on Ubuntu, then a reverse Meterpreter TCP session exfiltrates whole directories. Keylogging: SSH dictionary attack with Hydra, Metasploit session upgraded to sudo Meterpreter, logkeys records keystrokes, recordings downloaded. Table 5: keylogging 1,469, data theft 118 records.",
     "evidence": "Koroniotis 2019 PDF p.20: \"Post exploitation, we set-up a reverse meterpeter tcp connection through which exfiltration of entire directories became possible.\"",
     "techniques": [
      {
       "id": "T1041",
       "name": "Exfiltration Over C2 Channel",
       "tactics": "exfiltration",
       "fit": "primary"
      },
      {
       "id": "T1056.001",
       "name": "Keylogging",
       "tactics": "collection, credential-access",
       "fit": "secondary"
      },
      {
       "id": "T1110.001",
       "name": "Password Guessing",
       "tactics": "credential-access",
       "fit": "secondary"
      }
     ],
     "primary_tactic": "exfiltration",
     "confidence": "medium",
     "netflow_observable": "partial",
     "caveat": "Keylogging happens on the host and is invisible to NetFlow; only the Hydra SSH guessing, the Meterpreter session and the download are visible. By record count the label is about 93% keylogging-scenario traffic, so the flows may actually reflect credential-access/C2 rather than exfiltration. Sarhan Table 3 defines Theft only as 'data theft and keylogging' (1,909 flows). Very small class, a single scenario, EternalBlue (T1210) is also present but not listed to stay within three techniques."
    },
    {
     "label": "Benign",
     "documented_behaviour": "Normal background traffic generated constantly during the capture, alongside the attacks.",
     "evidence": "Koroniotis 2019 PDF p.17: \"normal background traffic being constantly generated\" (Section 4.1).",
     "techniques": [],
     "primary_tactic": "none (not adversary behaviour)",
     "confidence": "high",
     "netflow_observable": "yes",
     "caveat": "Benign has no ATT&CK technique. Benign flows were captured concurrently with attacks in the same lab, so a source-specific signature is likely (see the source probe)."
    }
   ]
  },
  "challenge": {
   "source": "NF-BoT-IoT-v3",
   "verdicts": [
    {
     "label": "DDoS",
     "mapping_ok": false,
     "problem": "All IDs exist, none deprecated or revoked, and the tactics match the STIX file. (1) T1499.003 is wrong for GoldenEye. The STIX text for T1499.002 Service Exhaustion Flood names the 'simple HTTP flood ... large number of HTTP requests to a web server' as its example. T1499.003 covers repeated requests to resource-intensive application features, which the paper does not document. Replace it with T1499.002. (2) The hping3 --syn --flood example is the textbook case in the T1499.001 STIX text ('TCP state-exhaustion attacks such as SYN floods'). TCP is about 51% of records and UDP about 49%, and only the SYN command is quoted. So T1498.001 as primary is a judgement call. The confidence is high for the tactic (impact) but only medium for the technique, and the single 'high' overstates it. (3) The quote is trimmed ('[54]' and 'for example,' dropped) but still fair, and p.19 and Table 5 on p.20 check out. (4) The HTTP/GoldenEye share is about 0.1% of records, so T1499.002 is only nominal for this class. NetFlow observability 'yes' is correct for the TCP/UDP floods; the HTTP part is only partially observable.",
     "corrected_techniques": "T1498.001 Direct Network Flood (primary, UDP and TCP volume); T1499.001 OS Exhaustion Flood (secondary, SYN flood); T1499.002 Service Exhaustion Flood (secondary, GoldenEye HTTP flood, replaces T1499.003)",
     "confidence_after": "medium"
    },
    {
     "label": "DoS",
     "mapping_ok": false,
     "problem": "(1) Same technique fix as DDoS: replace T1499.003 with T1499.002, because an HTTP flood is the T1499.002 example. (2) documented_behaviour says 'from a single attacker', but Koroniotis p.18 says the opposite: 'DDoS and DoS attacks are performed by a group of compromised machines called Bots'. The single-source claim is not in the paper. It only follows from the Sarhan Table 3 wording, which says DDoS has 'multiple different distributed sources'. State this as a dataset-level inference, not as paper evidence. (3) The evidence quote ('attempt to disrupt a service ...') is on PDF p.18, not p.19; only the hping3 and GoldenEye tool text is on p.19. (4) The paper does not support DoS being a separate technique at all, and the caveat already says this. Confidence is medium for the same reasons as DDoS. NetFlow 'yes' is fine.",
     "corrected_techniques": "T1498.001 Direct Network Flood (primary); T1499.001 OS Exhaustion Flood (secondary); T1499.002 Service Exhaustion Flood (secondary, replaces T1499.003)",
     "confidence_after": "medium"
    },
    {
     "label": "Reconnaissance",
     "mapping_ok": false,
     "problem": "T1595 as primary and T1046 as secondary are sound; both IDs exist and are valid. The recon quote and its p.18 citation verify. The error is in the OS-fingerprinting technique. In the STIX file, T1592.002 Software covers installed software types and versions. OS and version are listed under T1592.004 Client Configurations ('operating system/version, virtualization, architecture'). The caveat that T1592.002 'covers OS only loosely' hides that a better ID exists. Use T1592.004 for the Nmap -O and Xprobe2 part (about 20% of records). T1592.002 is justifiable only for the Nmap -sV service-version scans. The paper wrongly describes sV as 'a Syn scan', so the version-detection use is not clearly documented. T1595.001 Scanning IP Blocks is correctly excluded because single hosts were targeted. NetFlow 'yes' is right, but a flow cannot tell service scanning from OS fingerprinting. Confidence stays high for the label and tactic.",
     "corrected_techniques": "T1595 Active Scanning (primary, reconnaissance); T1046 Network Service Discovery (secondary, discovery); T1592.004 Client Configurations (secondary, OS fingerprinting; replaces T1592.002)",
     "confidence_after": "high"
    },
    {
     "label": "Theft",
     "mapping_ok": false,
     "problem": "(1) The evidence citation is wrong: the exfiltration quote is on PDF p.19, not p.20; only Table 5 is on p.20. (2) The quote is not verbatim. The paper spells 'meterpeter'; the mapping silently corrected it to 'meterpreter', and the quote also drops the line-break hyphen. (3) T1210 Exploitation of Remote Services was left out only to 'stay within three techniques'. That cap is arbitrary. The paper explicitly documents the SMB EternalBlue exploit on Windows 7, and the STIX tactic for T1210 is lateral-movement. Add it, and list the cap as a rule, not as evidence. (4) T1041 fits the data-theft sentence well (reverse Meterpreter TCP session, exfiltration of directories). The keylogging scenario downloads recordings through the same session, so T1041 covers that too. T1110.001 for Hydra and T1056.001 for logkeys are correct IDs with the right tactics. (5) The primary tactic 'exfiltration' is the paper-level semantics ('Information Theft'). The caveat that about 93% of Table 5 records are keylogging-scenario flows (credential-access and C2 traffic) means the flow-level tactic is uncertain. Also, Table 5 gives 1,587 records while Sarhan Table 3 gives 1,909 flows, so the 93% figure is from the original Bot-IoT records, not NF-BoT-IoT-v3. Medium confidence is acceptable if this is stated. NetFlow 'partial' is right.",
     "corrected_techniques": "T1041 Exfiltration Over C2 Channel (primary); T1110.001 Password Guessing (secondary, Hydra SSH); T1056.001 Keylogging (secondary, host-side, not NetFlow-visible); T1210 Exploitation of Remote Services (secondary, EternalBlue SMB)",
     "confidence_after": "medium"
    },
    {
     "label": "Benign",
     "mapping_ok": true,
     "problem": "No technique is correct, because benign traffic is not adversary behaviour. NetFlow 'yes' is fine. The quote 'normal background traffic being constantly generated' is verbatim but is on PDF p.13 (Section 4.1), not p.17, so fix the citation. The papers also give no benign flow composition (apps or protocols) for NF-BoT-IoT-v3. The Sarhan Table 3 benign count of 13,859 is for NF-BoT-IoT, not necessarily v3, and should be labelled that way. The source-signature caveat is a reasonable inference, not paper evidence.",
     "confidence_after": "high"
    }
   ]
  }
 },
 {
  "source": "NF-CSE-CIC-IDS2018-v3",
  "mapping": {
   "source": "NF-CSE-CIC-IDS2018-v3",
   "papers_read": [
    "2017_Sharafaldin_Towards_a_Reliable_Intrusion_Detection_Benchmark_Dataset.pdf (CIC-IDS2017 generator paper, NOT the 2018 dataset paper; the 2018 collection paper is not in References)",
    "2021_Sarhan_NetFlow_Datasets_for_Machine_Learning_Based_Network_Intrusio.pdf",
    "2021_Sarhan_Towards_a_Standard_Feature_Set_for_Network_Intrusion_Detecti.pdf",
    "2019_Ring_A_survey_of_network_based_intrusion_detection_data_sets.pdf (CICIDS2017 tool list only)",
    "Keyword-searched ~130 other PDFs in References for CSE-CIC-IDS2018 attack descriptions (HOIC, Slowloris, GoldenEye, Patator etc.); none describes the 2018 scenarios"
   ],
   "classes": [
    {
     "label": "Benign",
     "documented_behaviour": "Normal user traffic generated from abstracted human-user behaviour; not an attack.",
     "evidence": "Sarhan_2021_NetFlow_Datasets.pdf p.10 Table 5: Benign 'Normal unmalicious flows'.",
     "techniques": [],
     "primary_tactic": "none (no ATT&CK mapping)",
     "confidence": "high",
     "netflow_observable": "yes",
     "caveat": "Benign is the negative class; ATT&CK describes adversary behaviour only."
    },
    {
     "label": "Bot",
     "documented_behaviour": "Hijacked hosts remotely controlled to perform malicious activity. Papers read name no bot family or tool for 2018 (the 2017 survey lists Ares for CICIDS2017 only).",
     "evidence": "Sarhan_2021_NetFlow_Datasets.pdf p.10 Table 5: 'enables an attacker to remotely control several hijacked computers to perform malicious activities.'",
     "techniques": [
      {
       "id": "T1071",
       "name": "Application Layer Protocol",
       "tactics": "command-and-control",
       "fit": "primary"
      },
      {
       "id": "T1583.005",
       "name": "Botnet",
       "tactics": "resource-development",
       "fit": "secondary"
      }
     ],
     "primary_tactic": "command-and-control",
     "confidence": "medium",
     "netflow_observable": "partial",
     "caveat": "Sub-technique T1071.001 not chosen because the papers do not state HTTP C2 for 2018. Flows show periodic C2 beaconing at best; T1583.005 is adversary preparation and is not directly observable in flows."
    },
    {
     "label": "Brute_Force_-Web",
     "documented_behaviour": "Dictionary brute force against web service accounts. The 2018 scenario is not described in the papers read. The 2017 generator paper describes brute force against at least two services, and Sarhan lumps Brute Force-Web into the BruteForce parent class.",
     "evidence": "Sharafaldin_2017 p.15: 'acquiring different services accounts by running a dictionary brute force attack against the main server (at least two major services...)'. Sarhan_2021 p.14: Brute Force-Web merged into brute-force.",
     "techniques": [
      {
       "id": "T1110.001",
       "name": "Password Guessing",
       "tactics": "credential-access",
       "fit": "primary"
      }
     ],
     "primary_tactic": "credential-access",
     "confidence": "medium",
     "netflow_observable": "partial",
     "caveat": "Per-label 2018 behaviour inferred from label name plus the 2017 generator description. Many HTTP login attempts appear as repeated flows to one server, but flows cannot show credentials or success."
    },
    {
     "label": "Brute_Force_-XSS",
     "documented_behaviour": "Cross-site scripting: attacker uses web applications to send malicious scripts to end users. The 2018 scenario details are not in the papers read. The 2017 web-attack scenario scans a vulnerable web app, then runs SQLi, command injection and file upload; XSS is only a label in Sarhan.",
     "evidence": "Sarhan_2021_NetFlow_Datasets.pdf p.9-10 Table 6 (XSS definition): 'Cross-site Scripting is a type of injection in which an attacker uses web applications to send malicious scripts to end-users.'",
     "techniques": [
      {
       "id": "T1190",
       "name": "Exploit Public-Facing Application",
       "tactics": "initial-access",
       "fit": "primary"
      },
      {
       "id": "T1059.007",
       "name": "JavaScript",
       "tactics": "execution",
       "fit": "secondary"
      }
     ],
     "primary_tactic": "initial-access",
     "confidence": "low",
     "netflow_observable": "no",
     "caveat": "Label name says 'Brute Force' but Sarhan defines XSS as injection; Sarhan merges it into brute-force in NF-UQ-NIDS, so the label is ambiguous and may be a web brute-force/XSS mix. Payloads are in HTTP content, invisible to NetFlow; only repeated HTTP requests to one host are visible. Earlier dedup found 66 distinct vectors for this label, so it is nearly uninformative."
    },
    {
     "label": "DDOS_attack-HOIC",
     "documented_behaviour": "Distributed DoS flood; only the parent DDoS definition is given. The tool (High Orbit Ion Cannon, HTTP flood) is known from the label and not described in the papers read.",
     "evidence": "Sarhan_2021_NetFlow_Datasets.pdf p.10 Table 5: DDoS 'An attempt similar to DoS but has multiple different distributed sources.' Renaming to DDoS on p.8.",
     "techniques": [
      {
       "id": "T1499.002",
       "name": "Service Exhaustion Flood",
       "tactics": "impact",
       "fit": "primary"
      },
      {
       "id": "T1498.001",
       "name": "Direct Network Flood",
       "tactics": "impact",
       "fit": "secondary"
      }
     ],
     "primary_tactic": "impact",
     "confidence": "low",
     "netflow_observable": "yes",
     "caveat": "HTTP-layer flood assumed from the tool name, not from the papers. ATT&CK has no distributed-source field; DDoS vs DoS is not an ATT&CK distinction, so source count must be derived from flows."
    },
    {
     "label": "DDOS_attack-LOIC-UDP",
     "documented_behaviour": "Distributed UDP flood with LOIC; the papers only give the parent DDoS definition. The Ring survey says LOIC was used for DDoS in CICIDS2017, not 2018.",
     "evidence": "Sarhan_2021_NetFlow_Datasets.pdf p.10 Table 5: DDoS 'similar to DoS but has multiple different distributed sources.' Ring_2019 p.8 lists 'DDoS (executed through LOIC)' for CICIDS2017.",
     "techniques": [
      {
       "id": "T1498.001",
       "name": "Direct Network Flood",
       "tactics": "impact",
       "fit": "primary"
      }
     ],
     "primary_tactic": "impact",
     "confidence": "medium",
     "netflow_observable": "yes",
     "caveat": "UDP transport comes from the label name only; the 2018 scenario is not described in the papers read. Volumetric floods are well captured in flows."
    },
    {
     "label": "DDoS_attacks-LOIC-HTTP",
     "documented_behaviour": "Distributed HTTP flood with LOIC; the papers only give the parent DDoS definition and (for 2017) LOIC as the DDoS tool.",
     "evidence": "Ring_2019 p.8: CICIDS2017 'DDoS (executed through LOIC)'. Sarhan_2021 p.10 Table 5: DDoS 'multiple different distributed sources.'",
     "techniques": [
      {
       "id": "T1499.002",
       "name": "Service Exhaustion Flood",
       "tactics": "impact",
       "fit": "primary"
      },
      {
       "id": "T1498.001",
       "name": "Direct Network Flood",
       "tactics": "impact",
       "fit": "secondary"
      }
     ],
     "primary_tactic": "impact",
     "confidence": "medium",
     "netflow_observable": "yes",
     "caveat": "Evidence is about 2017, not 2018, and the HTTP layer comes from the label name. Per-request HTTP content is not in NetFlow; only flow counts and volumes."
    },
    {
     "label": "DoS_attacks-GoldenEye",
     "documented_behaviour": "HTTP DoS tool. The 2018 scenario is not described in the papers read. The 2017 generator paper describes HTTP DoS tools that hold full TCP connections open with valid incomplete HTTP requests.",
     "evidence": "Ring_2019 p.8: CICIDS2017 'DoS (executed through Hulk, GoldenEye, Slowloris, and Slowhttptest)'. Sharafaldin_2017 p.15: tools 'start by making a full TCP connection... holds the connection open'.",
     "techniques": [
      {
       "id": "T1499.002",
       "name": "Service Exhaustion Flood",
       "tactics": "impact",
       "fit": "primary"
      }
     ],
     "primary_tactic": "impact",
     "confidence": "medium",
     "netflow_observable": "partial",
     "caveat": "Evidence concerns CIC-IDS2017, not 2018. GoldenEye issues keep-alive HTTP requests; this is inferred from tool knowledge, not the papers. The attack is a single-source DoS."
    },
    {
     "label": "DoS_attacks-Hulk",
     "documented_behaviour": "HTTP flood DoS tool. The papers read name Hulk only as a CICIDS2017 DoS tool and do not describe its behaviour.",
     "evidence": "Ring_2019 p.8: CICIDS2017 'DoS (executed through Hulk, GoldenEye, Slowloris, and Slowhttptest)'.",
     "techniques": [
      {
       "id": "T1499.002",
       "name": "Service Exhaustion Flood",
       "tactics": "impact",
       "fit": "primary"
      }
     ],
     "primary_tactic": "impact",
     "confidence": "medium",
     "netflow_observable": "partial",
     "caveat": "Hulk behaviour (many unique HTTP requests) is from tool knowledge, not the papers read, and the evidence is about 2017. Mapping to T1499.002 follows the ATT&CK text on simple HTTP floods."
    },
    {
     "label": "DoS_attacks-SlowHTTPTest",
     "documented_behaviour": "Low-and-slow HTTP DoS holding connections open with valid but incomplete HTTP requests until the server's sockets are exhausted. The generator paper describes this class of tool in general and does not name SlowHTTPTest for 2018.",
     "evidence": "Sharafaldin_2017 p.15: 'holds the connection open by sending valid, incomplete HTTP requests... all sockets are used up and no other connection can be made.'",
     "techniques": [
      {
       "id": "T1499.002",
       "name": "Service Exhaustion Flood",
       "tactics": "impact",
       "fit": "primary"
      }
     ],
     "primary_tactic": "impact",
     "confidence": "medium",
     "netflow_observable": "partial",
     "caveat": "ATT&CK has no slow-DoS sub-technique; T1499.002 is the closest fit but describes volume floods. The paper is the 2017 generator model, not the 2018 dataset. Earlier dedup found this label collapses to a single distinct feature vector."
    },
    {
     "label": "DoS_attacks-Slowloris",
     "documented_behaviour": "Same slow-HTTP DoS mechanism: full TCP connections kept open with incomplete HTTP requests, exhausting server sockets.",
     "evidence": "Sharafaldin_2017 p.15: 'sending valid, incomplete HTTP requests to the server at regular intervals to keep the sockets from closing.' Ring_2019 p.8 lists Slowloris for CICIDS2017.",
     "techniques": [
      {
       "id": "T1499.002",
       "name": "Service Exhaustion Flood",
       "tactics": "impact",
       "fit": "primary"
      }
     ],
     "primary_tactic": "impact",
     "confidence": "medium",
     "netflow_observable": "partial",
     "caveat": "Same caveats as SlowHTTPTest: the 2017 generator description is applied to 2018, and no slow-DoS sub-technique exists. Long-lived low-rate flows are visible, but the incomplete HTTP requests are not."
    },
    {
     "label": "FTP-BruteForce",
     "documented_behaviour": "Dictionary password guessing against an FTP service. Sarhan merges it into BruteForce, defined as obtaining credentials from a list of predefined possibilities. The FTP tool is not named in the papers read.",
     "evidence": "Sarhan_2021_NetFlow_Datasets.pdf p.10 Table 5: BruteForce 'obtain usernames and password credentials by accessing a list of predefined possibilities.' Sharafaldin_2017 p.15: dictionary brute force.",
     "techniques": [
      {
       "id": "T1110.001",
       "name": "Password Guessing",
       "tactics": "credential-access",
       "fit": "primary"
      }
     ],
     "primary_tactic": "credential-access",
     "confidence": "high",
     "netflow_observable": "partial",
     "caveat": "Repeated short FTP control connections from one source are visible, but credentials and success are not. The label collapses to a single distinct vector after dedup, so it is a fingerprint of one tool run."
    },
    {
     "label": "SSH-Bruteforce",
     "documented_behaviour": "Dictionary password guessing against SSH. The 2017 generator paper lists SSH password brute force; the Ring survey lists SSH brute force for CICIDS2017.",
     "evidence": "Ring_2019 p.8: CICIDS2017 'heartbleed, infiltration, SSH brute force, SQL injection'. Sarhan_2021 p.10 Table 5: BruteForce 'list of predefined possibilities.'",
     "techniques": [
      {
       "id": "T1110.001",
       "name": "Password Guessing",
       "tactics": "credential-access",
       "fit": "primary"
      }
     ],
     "primary_tactic": "credential-access",
     "confidence": "high",
     "netflow_observable": "partial",
     "caveat": "Repeated SSH sessions from one source are visible; encrypted content hides credentials and success. Evidence for SSH is about 2017. The successful-login follow-up (T1021.004) is not in this label."
    },
    {
     "label": "Infilteration",
     "documented_behaviour": "Inside attack: a malicious file is emailed to a victim, exploiting an application, then a backdoor scans the network for other vulnerable hosts. The 2017 generator paper describes a vulnerable document viewer exploited from an email.",
     "evidence": "Sarhan_2021_NetFlow_Datasets.pdf p.10 Table 5: 'inside attack that sends a malicious file via an email to exploit an application and is followed by a backdoor that scans the network.'",
     "techniques": [
      {
       "id": "T1566.001",
       "name": "Spearphishing Attachment",
       "tactics": "initial-access",
       "fit": "primary"
      },
      {
       "id": "T1203",
       "name": "Exploitation for Client Execution",
       "tactics": "execution",
       "fit": "secondary"
      },
      {
       "id": "T1046",
       "name": "Network Service Discovery",
       "tactics": "discovery",
       "fit": "secondary"
      }
     ],
     "primary_tactic": "initial-access",
     "confidence": "medium",
     "netflow_observable": "partial",
     "caveat": "Multi-stage: the email and exploit are not in flows; only the post-compromise internal scanning (T1046) and backdoor C2 are, so the labelled flows probably capture the discovery stage, not the primary technique. Sarhan gives a summary, not a per-stage description. Label is spelled 'Infilteration' in the data."
    },
    {
     "label": "SQL_Injection",
     "documented_behaviour": "Injection of untrusted input into a web application to alter query execution. The 2017 generator paper lists SQL injection in its web-attack scenario after a vulnerability scan.",
     "evidence": "Sharafaldin_2017 p.15: vulnerable web app, 'different web attacks... (including SQL injection, command injection, unrestricted file upload).' Sarhan_2021 Table 6: injection definition.",
     "techniques": [
      {
       "id": "T1190",
       "name": "Exploit Public-Facing Application",
       "tactics": "initial-access",
       "fit": "primary"
      }
     ],
     "primary_tactic": "initial-access",
     "confidence": "high",
     "netflow_observable": "no",
     "caveat": "The payload is in HTTP content, so NetFlow shows only HTTP requests to the web server, indistinguishable from benign browsing except by volume and timing. The scenario is described for 2017; 2018 is inferred. T1190 is the most specific supported ID; no SQLi sub-technique exists."
    }
   ]
  },
  "challenge": {
   "source": "NF-CSE-CIC-IDS2018-v3",
   "verdicts": [
    {
     "label": "Benign",
     "mapping_ok": true,
     "problem": "No technique needed. Citation fix only: the Benign row ('Normal unmalicious flows') is in Sarhan_2021_NetFlow_Datasets Table 5 on p.9, not p.10 (p.10 starts with Table 6, the NF-UQ-NIDS counts, which has no definitions). The same p.10 -> p.9 slip recurs in nearly every row below. Observability 'yes' is fine.",
     "confidence_after": "high"
    },
    {
     "label": "Bot",
     "mapping_ok": false,
     "problem": "T1583.005 is the wrong Botnet technique. In the STIX file T1583.005 (Acquire Infrastructure: Botnet, resource-development) means buying, leasing or renting a botnet. T1584.005 (Compromise Infrastructure: Botnet, resource-development) covers building a botnet by compromising third-party systems, which matches the paper's 'hijacked computers'. Neither is observable in flows. T1071 (command-and-control) as primary is acceptable, but it is the parent and the papers give no protocol. The quote is accurate; the page is p.9, not p.10. Confidence medium is acceptable because the quote is the 2018 class description. Observability 'partial' is fine.",
     "corrected_primary_tactic": "command-and-control",
     "corrected_techniques": "T1071 Application Layer Protocol (primary, command-and-control); T1584.005 Botnet (secondary, resource-development; replaces T1583.005)",
     "confidence_after": "medium"
    },
    {
     "label": "Brute_Force_-Web",
     "mapping_ok": false,
     "problem": "Evidence is misattributed. Sarhan_2021 p.14 does not say Brute Force-Web was merged into brute-force. That merge is in the NF-UQ-NIDS paragraph on p.8-9, and p.14 is Table 12 results. In the per-dataset NF-CSE-CIC-IDS2018 table (Table 5, p.9), Brute Force-Web is not in BruteForce. The counts show this. UQ Brute Force 291,955 minus the 2018 BruteForce 287,597 leaves 4,358. UQ Injection 468,575 minus ToN Injection 468,539 leaves 36. Together that is 4,394, the 2018 Web Attacks count. The v2 numbers also tie out: 3,070 + 432 = 3,502. So Sarhan's own description of these flows is the 'Web Attacks' group ('SQL injections, command injections and unrestricted file uploads'), which never mentions brute force. The 2018 scenario is not described in the papers read, and the Sharafaldin 2017 brute-force scenario is generic. The mapping comes from the label name alone, so the rule requires low, not medium. T1110.001 itself exists, is not deprecated, and is plausible. Observability 'partial' is fine; Sarhan Table 12 p.14 shows detection rates of about 50%.",
     "corrected_primary_tactic": "credential-access",
     "corrected_techniques": "T1110.001 Password Guessing (primary, credential-access); keep, but evidence = 'not described in the papers read' for 2018",
     "confidence_after": "low"
    },
    {
     "label": "Brute_Force_-XSS",
     "mapping_ok": false,
     "problem": "The quote is from the wrong dataset. 'Cross-site Scripting is a type of injection...' is the XSS row of Table 4 (NF-ToN-IoT, Sarhan_2021 p.9). It is not 'Table 6' and not a 2018 definition; the 2018 tables have no XSS class. The label is 'Brute Force -XSS', which mixes a credential-guessing name with an injection name. Sarhan puts these flows in the 2018 'Web Attacks' group (see the arithmetic under Brute_Force_-Web) and in the brute-force class in NF-UQ-NIDS. The papers that do mention it: Ring_2019 p.8 lists 'cross-site-scripting' for CICIDS2017, and Sarhan Table 12 p.14 lists the label with a detection rate of 49%. Neither describes the 2018 behaviour. ATT&CK has no XSS technique, because XSS is a weakness and not a behaviour. T1190 attacks the server rather than script delivery to users, and T1059.007 would run on the victim's browser. Both are weak, so the mapping is exploratory. If a brute-force reading is intended, T1110.001 is the alternative. Confidence low is right; netflow 'no' for the payload is right.",
     "corrected_primary_tactic": "initial-access",
     "corrected_techniques": "T1190 Exploit Public-Facing Application (primary, initial-access, weak fit); T1059.007 JavaScript (secondary, execution, weak fit); evidence = 'not described in the papers read' for 2018. Replace the ToN-IoT quote with Ring_2019 p.8 (CICIDS2017 lists cross-site-scripting).",
     "confidence_after": "low"
    },
    {
     "label": "DDOS_attack-HOIC",
     "mapping_ok": false,
     "problem": "Technique choice is defensible. T1499.002 text explicitly describes a 'simple HTTP flood', and T1498.001 as secondary is optional. In ATT&CK, T1499 is DoS without saturating the network, while T1498 saturates it. Observability 'yes' is overstated. The HTTP content is invisible, the caveat admits the HTTP layer is assumed, and Sarhan Table 12 p.14 shows only 45.66% DR for HOIC in NetFlow, with the text saying HOIC samples were 'mostly undetected'. It should be 'partial'. Citation fix: Table 5 is on p.9. Confidence low is right.",
     "corrected_primary_tactic": "impact",
     "corrected_techniques": "T1499.002 Service Exhaustion Flood (primary, impact); T1498.001 Direct Network Flood (secondary, optional); netflow_observable = partial",
     "confidence_after": "low"
    },
    {
     "label": "DDOS_attack-LOIC-UDP",
     "mapping_ok": false,
     "problem": "Technique T1498.001 is correct, and its text names UDP as a typical flood protocol. Confidence medium is too high under the rule, because UDP and LOIC come from the label name only. The papers give only the parent DDoS definition (Table 5, p.9), and Ring p.8 is about CICIDS2017 LOIC. It should be low. Observability 'yes' is acceptable for a volumetric flood, though Sarhan Table 12 p.14 shows only 81% DR. Citation fix: p.9, not p.10.",
     "corrected_primary_tactic": "impact",
     "corrected_techniques": "T1498.001 Direct Network Flood (primary, impact); keep",
     "confidence_after": "low"
    },
    {
     "label": "DDoS_attacks-LOIC-HTTP",
     "mapping_ok": false,
     "problem": "Confidence medium is inconsistent with the HOIC row, which has the same evidence quality (parent DDoS definition plus 2017 LOIC), and should be low. Observability 'yes' contradicts the row's own caveat that per-request HTTP content is not in NetFlow, so it should be 'partial'. Table 12 p.14 gives F1 0.71 for this label. T1499.002 primary is fine for an HTTP flood. T1498.001 as secondary is optional, since ATT&CK separates network saturation (T1498) from endpoint DoS (T1499). Citation fix: p.9.",
     "corrected_primary_tactic": "impact",
     "corrected_techniques": "T1499.002 Service Exhaustion Flood (primary, impact); T1498.001 Direct Network Flood (secondary, optional); netflow_observable = partial",
     "confidence_after": "low"
    },
    {
     "label": "DoS_attacks-GoldenEye",
     "mapping_ok": false,
     "problem": "The Sharafaldin p.15 quote does not name GoldenEye. It describes slow-HTTP tools that hold sockets open, and the caveat then says GoldenEye is a keep-alive HTTP flood, so the cited quote supports a different mechanism. Ring p.8 only lists GoldenEye as a CICIDS2017 tool. The 2018 class is not described in the papers read and the behaviour comes from tool knowledge, so the rule requires low, not medium. T1499.002 is an acceptable fit for an HTTP request flood. Observability 'partial' is fine; Table 12 p.14 shows 99.3% DR.",
     "corrected_primary_tactic": "impact",
     "corrected_techniques": "T1499.002 Service Exhaustion Flood (primary, impact); keep; evidence = 'not described in the papers read' for 2018",
     "confidence_after": "low"
    },
    {
     "label": "DoS_attacks-Hulk",
     "mapping_ok": false,
     "problem": "The only evidence is a tool name in a 2017 list (Ring p.8). The papers give no behaviour for Hulk in either 2017 or 2018, so confidence medium is wrong and the rule requires low. T1499.002 matches ATT&CK's 'simple HTTP flood' text and is a reasonable choice. The caveat already admits the behaviour comes from tool knowledge.",
     "corrected_primary_tactic": "impact",
     "corrected_techniques": "T1499.002 Service Exhaustion Flood (primary, impact); keep; evidence = 'not described in the papers read'",
     "confidence_after": "low"
    },
    {
     "label": "DoS_attacks-SlowHTTPTest",
     "mapping_ok": false,
     "problem": "T1499.002 does not fit. Its STIX text describes a 'simple HTTP flood' that 'relies on raw volume' (or SSL renegotiation), whereas slow-HTTP attacks are low-rate connection holding. The row's own caveat concedes this. T1499 (Endpoint Denial of Service) says it covers exhausting the hosting system's resources and has no volume requirement. It is the better primary, with T1499.002 as the closest sub-technique. Confidence medium is too high. The Sharafaldin p.15 passage describes a generic tool class for the 2017 generator and never names SlowHTTPTest, so the 2018 class is not described and the rule requires low. Observability 'partial' is arguably optimistic: Sarhan Table 12 p.14 reports 0.00% DR in NetFlow and the text says 'fully undetected'.",
     "corrected_primary_tactic": "impact",
     "corrected_techniques": "T1499 Endpoint Denial of Service (primary, impact; no sub-technique fits slow-DoS); T1499.002 Service Exhaustion Flood (secondary, closest sub-technique)",
     "confidence_after": "low"
    },
    {
     "label": "DoS_attacks-Slowloris",
     "mapping_ok": false,
     "problem": "Same issue as SlowHTTPTest. T1499.002 describes a volume-based HTTP flood, but Slowloris holds sockets open with incomplete requests at low rate, so T1499 is the better primary. The Sharafaldin p.15 quote matches the Slowloris mechanism well, but the paper never names Slowloris. Only Ring p.8 names it, for CICIDS2017, and the 2018 class is not described. Confidence should be low. Observability 'partial' is fine: Table 12 p.14 shows 99.95% DR, but that may reflect a tool signature and not the attack's flow behaviour.",
     "corrected_primary_tactic": "impact",
     "corrected_techniques": "T1499 Endpoint Denial of Service (primary, impact); T1499.002 Service Exhaustion Flood (secondary, closest sub-technique)",
     "confidence_after": "low"
    },
    {
     "label": "FTP-BruteForce",
     "mapping_ok": true,
     "problem": "T1110.001 is correct, and the STIX text lists FTP (21/TCP) as a commonly guessed service. Sarhan Table 5 'BruteForce' (p.9, not p.10) is a direct description of the 2018 class. The counts show Table 5 BruteForce (287,597) is FTP plus SSH, since Web and XSS are the 4,358 in 'Web Attacks'. High confidence is defensible. The Sharafaldin 'dictionary brute force' quote is accurate (p.15). Observability 'partial' is fine; Table 12 shows 100% DR, but the earlier dedup finding suggests a tool fingerprint.",
     "corrected_primary_tactic": "credential-access",
     "corrected_techniques": "T1110.001 Password Guessing (primary, credential-access); keep",
     "confidence_after": "high"
    },
    {
     "label": "SSH-Bruteforce",
     "mapping_ok": true,
     "problem": "T1110.001 is correct, and SSH (22/TCP) is listed in the STIX text. The Ring p.8 quote is accurate but about CICIDS2017. The direct 2018 support is Sarhan Table 5 BruteForce, p.9 and not p.10, which equals FTP+SSH by the count arithmetic. The first agent should cite that as primary evidence. High confidence is acceptable on that basis. Observability 'partial' is fine, and Table 12 shows 99.93% DR.",
     "corrected_primary_tactic": "credential-access",
     "corrected_techniques": "T1110.001 Password Guessing (primary, credential-access); keep",
     "confidence_after": "high"
    },
    {
     "label": "Infilteration",
     "mapping_ok": true,
     "problem": "The mapping holds. T1566.001, T1203 and T1046 exist, are not deprecated, and have the stated tactics. The Sarhan Table 5 quote is accurate (p.9, not p.10), and it is a direct description of the 2018 class. A stronger second source was missed: Sharafaldin_2017 p.14 describes the 'vulnerable document viewer', a malicious document received by email, a backdoor, and an internal scan for other vulnerable boxes. That matches T1566.001, T1203 and T1046 closely. Confidence medium is right, because the 2018 paper is absent and the 2017 scenario is a design, not a result. The caveat is correct that only discovery or C2 traffic is visible in flows. Table 12 p.14 shows 62.7% DR but F1 0.04. Keeping initial-access as primary tactic is defensible because the label describes the whole chain; the observable stage is discovery.",
     "corrected_primary_tactic": "initial-access",
     "corrected_techniques": "T1566.001 Spearphishing Attachment (primary, initial-access); T1203 Exploitation for Client Execution (secondary, execution); T1046 Network Service Discovery (secondary, discovery); add Sharafaldin_2017 p.14 as evidence",
     "confidence_after": "medium"
    }
   ]
  }
 }
]