\section{Evaluability of every label under P1 and P2}\label{app:eval-p12} Table~\ref{tab:app-eval-p12} gives, for every label, the number of training and test rows, the number of clean test rows (feature set B) and whether the label passes the thresholds for feature sets A / B (y = yes, n = no). ``Distinct B'' counts distinct feature-set-B vectors among the P2 training rows. {\tiny \setlength{\tabcolsep}{2pt} \begin{longtable}{lrrrcrrrrc} \caption{Training and test rows per label under P1 and P2.}\label{tab:app-eval-p12}\\ \toprule Label & P1 train & P1 test & P1 clean & A/B & P2 train & P2 test & P2 clean & Distinct B & A/B \\ \midrule \endfirsthead \toprule Label & P1 train & P1 test & P1 clean & A/B & P2 train & P2 test & P2 clean & Distinct B & A/B \\ \midrule \endhead \bottomrule \endfoot Benign & 24\,322\,329 & 5\,211\,931 & 2\,380\,127 & y/y & 24\,243\,652 & 5\,156\,672 & 3\,084\,727 & 11\,454\,409 & y/y \\ DDoS & 7\,904\,496 & 1\,693\,822 & 80\,388 & y/y & 7\,663\,141 & 1\,622\,230 & 1\,059\,650 & 615\,827 & y/y \\ DoS & 5\,770\,531 & 1\,236\,544 & 32\,274 & y/y & 5\,607\,021 & 1\,035\,116 & 995\,628 & 341\,154 & y/y \\ xss & 1\,984\,104 & 425\,166 & 375\,042 & y/y & 1\,953\,067 & 394\,307 & 356\,663 & 1\,801\,678 & y/y \\ Reconnaissance & 1\,198\,543 & 256\,832 & 5\,349 & y/y & 1\,197\,758 & 92\,332 & 4\,174 & 29\,397 & y/y \\ password & 1\,116\,343 & 239\,217 & 97\,327 & y/y & 1\,099\,244 & 227\,491 & 96\,823 & 507\,889 & y/y \\ scanning & 951\,283 & 203\,847 & 5\,694 & y/y & 905\,605 & 899 & 764 & 31\,650 & y/y \\ DDOS\_attack-HOIC & 722\,617 & 154\,847 & 27\,666 & y/y & 712\,001 & 146\,501 & 83\,932 & 121\,736 & y/y \\ injection & 267\,243 & 57\,267 & 51\,767 & y/y & 262\,396 & 52\,037 & 46\,813 & 238\,504 & y/y \\ DDoS\_attacks-LOIC-HTTP & 202\,012 & 43\,289 & 39\,206 & y/y & 192\,886 & 42\,114 & 39\,058 & 178\,428 & y/y \\ FTP-BruteForce & 135\,352 & 29\,004 & 0 & y/n & 133\,430 & 27\,060 & 0 & 1 & y/n \\ Infilteration & 81\,063 & 17\,371 & 11\,196 & y/y & 79\,462 & 13\,726 & 11\,119 & 56\,844 & y/y \\ Bot & 78\,022 & 16\,719 & 324 & y/y & 75\,072 & 13\,771 & 708 & 1\,817 & y/y \\ Backdoor & 74\,444 & 15\,953 & 3\,102 & y/y & 74\,043 & 14\,874 & 2\,171 & 17\,959 & y/y \\ DoS\_attacks-SlowHTTPTest & 73\,885 & 15\,833 & 0 & y/n & 72\,830 & 14\,780 & 0 & 1 & y/n \\ DoS\_attacks-Hulk & 70\,053 & 15\,012 & 15\,010 & y/y & 68\,844 & 13\,679 & 13\,678 & 68\,841 & y/y \\ SSH-Bruteforce & 65\,965 & 14\,136 & 14\,084 & y/y & 65\,019 & 13\,196 & 13\,159 & 64\,893 & y/y \\ Exploits & 29\,920 & 6\,412 & 5\,970 & y/y & 29\,666 & 6\,152 & 5\,817 & 28\,176 & y/y \\ Fuzzers & 23\,671 & 5\,073 & 3\,904 & y/y & 23\,602 & 4\,996 & 3\,749 & 19\,424 & y/y \\ DoS\_attacks-GoldenEye & 21\,455 & 4\,598 & 4\,089 & y/y & 20\,470 & 3\,652 & 2\,548 & 18\,687 & y/y \\ Generic & 13\,755 & 2\,948 & 804 & y/y & 13\,641 & 2\,776 & 754 & 5\,991 & y/y \\ DoS\_attacks-Slowloris & 12\,614 & 2\,703 & 407 & y/y & 12\,437 & 2\,531 & 1\,151 & 1\,714 & y/y \\ mitm & 3\,687 & 791 & 727 & y/y & 3\,581 & 775 & 775 & 3\,321 & y/y \\ ransomware & 2\,779 & 596 & 399 & y/y & 2\,532 & 593 & 521 & 1\,850 & y/y \\ Shellcode & 1\,666 & 358 & 252 & y/y & 1\,658 & 340 & 188 & 1\,262 & y/y \\ DDOS\_attack-LOIC-UDP & 1\,207 & 259 & 259 & y/y & 1\,200 & 235 & 235 & 1\,200 & y/y \\ Brute\_Force\_-Web & 1\,132 & 243 & 79 & y/y & 1\,114 & 229 & 74 & 620 & y/y \\ Theft & 1\,130 & 243 & 47 & y/n & 794 & 0 & 0 & 334 & n/n \\ Analysis & 858 & 184 & 184 & y/y & 805 & 140 & 140 & 805 & y/y \\ Brute\_Force\_-XSS & 322 & 69 & 3 & y/n & 313 & 66 & 0 & 59 & y/n \\ SQL\_Injection & 308 & 66 & 11 & y/n & 304 & 63 & 20 & 94 & y/n \\ Worms & 110 & 24 & 23 & n/n & 110 & 23 & 20 & 98 & n/n \\ \end{longtable}} \section{Source-address statistics and evaluability under P3}\label{app:eval-p3} Table~\ref{tab:app-eval-p3} lists, for every label, the number of distinct source addresses (per source dataset), the addresses assigned to the training and test partitions, the share of the label's flows from its most frequent address, and the resulting numbers of rows. {\tiny \setlength{\tabcolsep}{2pt} \begin{longtable}{lrrrrrrc} \caption{Source addresses and rows per label under P3.}\label{tab:app-eval-p3}\\ \toprule Label & Addr. & Train addr. & Test addr. & Top share & Train rows & Test rows & Eval. \\ \midrule \endfirsthead \toprule Label & Addr. & Train addr. & Test addr. & Top share & Train rows & Test rows & Eval. \\ \midrule \endhead \bottomrule \endfoot DoS\_attacks-Hulk & 1 & 1 & 0 & 0.000 & 100\,076 & 0 & no \\ FTP-BruteForce & 1 & 1 & 0 & 0.000 & 193\,360 & 0 & no \\ DoS\_attacks-SlowHTTPTest & 1 & 0 & 0 & 0.000 & 0 & 0 & no \\ DoS\_attacks-GoldenEye & 1 & 1 & 0 & 0.000 & 30\,650 & 0 & no \\ SSH-Bruteforce & 1 & 1 & 0 & 0.000 & 94\,237 & 0 & no \\ DoS\_attacks-Slowloris & 1 & 0 & 1 & 0.000 & 0 & 18\,020 & no \\ mitm & 2 & 0 & 0 & 0.000 & 0 & 0 & no \\ ransomware & 3 & 1 & 2 & 0.000 & 1\,482 & 2\,489 & no \\ Exploits & 4 & 4 & 0 & 0.000 & 42\,744 & 0 & no \\ Brute\_Force\_-XSS & 4 & 1 & 3 & 0.003 & 199 & 261 & no \\ Fuzzers & 4 & 4 & 0 & 0.000 & 33\,816 & 0 & no \\ Theft & 4 & 3 & 1 & 0.001 & 1\,382 & 233 & no \\ Generic & 4 & 4 & 0 & 0.000 & 19\,651 & 0 & no \\ Shellcode & 4 & 4 & 0 & 0.001 & 2\,381 & 0 & no \\ Analysis & 4 & 4 & 0 & 0.001 & 1\,226 & 0 & no \\ Worms & 4 & 4 & 0 & 0.007 & 158 & 0 & no \\ SQL\_Injection & 4 & 2 & 2 & 0.002 & 231 & 209 & no \\ xss & 5 & 3 & 1 & 0.000 & 1\,326\,340 & 242\,712 & no \\ password & 7 & 4 & 1 & 0.000 & 1\,154\,941 & 255\,565 & no \\ Backdoor & 8 & 6 & 2 & 0.000 & 4\,752 & 101\,598 & no \\ injection & 8 & 4 & 2 & 0.000 & 68\,333 & 259\,011 & no \\ Bot & 8 & 7 & 1 & 0.000 & 96\,243 & 15\,217 & no \\ DDOS\_attack-LOIC-UDP & 10 & 6 & 1 & 0.001 & 1\,045 & 170 & no \\ DDoS\_attacks-LOIC-HTTP & 10 & 6 & 1 & 0.000 & 174\,707 & 27\,436 & no \\ DDOS\_attack-HOIC & 10 & 6 & 1 & 0.000 & 619\,061 & 103\,588 & no \\ Reconnaissance & 12 & 10 & 1 & 0.000 & 1\,289\,341 & 422\,577 & no \\ DDoS & 14 & 8 & 2 & 0.000 & 7\,541\,946 & 1\,779\,074 & no \\ scanning & 17 & 11 & 4 & 0.000 & 914\,777 & 4\,731 & no \\ DoS & 19 & 14 & 1 & 0.000 & 6\,146\,142 & 2\,047\,744 & no \\ Brute\_Force\_-Web & 561 & 383 & 87 & 0.002 & 1\,067 & 334 & yes \\ Infilteration & 609 & 445 & 76 & 0.000 & 85\,358 & 156 & yes \\ Benign & 197\,005 & 137\,596 & 29\,845 & 0.000 & 23\,003\,891 & 3\,121\,123 & yes \\ \end{longtable}} \section{Leave-one-source-out: shared classes}\label{app:eval-p4} Table~\ref{tab:app-eval-p4} lists the labels that pass the thresholds for each held-out source, with the training rows (all other sources), the test rows (held-out source), the test rows that duplicate a training vector with the same label (``Same) or only with another label (``Other), the clean test rows, the distinct feature-set-B training vectors (``Dist.\ B) and whether the label is evaluable for feature set B. {\tiny \setlength{\tabcolsep}{2pt} \begin{longtable}{llrrrrrrc} \caption{Leave-one-source-out, shared classes: rows per evaluable label.}\label{tab:app-eval-p4}\\ \toprule Held out & Label & Train & Test & Same & Other & Clean & Dist.\ B & Eval B \\ \midrule \endfirsthead \toprule Held out & Label & Train & Test & Same & Other & Clean & Dist.\ B & Eval B \\ \midrule \endhead \bottomrule \endfoot UNSW-NB15 & Backdoor & 91\,466 & 4\,658 & 0 & 0 & 4\,658 & 18\,409 & yes \\ UNSW-NB15 & Benign & 29\,274\,486 & 2\,222\,930 & 1 & 0 & 2\,222\,929 & 13\,840\,904 & yes \\ UNSW-NB15 & DoS & 7\,414\,219 & 5\,971 & 0 & 0 & 5\,971 & 458\,738 & yes \\ UNSW-NB15 & Reconnaissance & 1\,525\,740 & 17\,074 & 0 & 0 & 17\,074 & 26\,031 & yes \\ ToN-IoT & Backdoor & 4\,217 & 101\,692 & 0 & 2\,842 & 98\,850 & 3\,522 & yes \\ ToN-IoT & Benign & 17\,703\,512 & 15\,078\,514 & 5\,303\,956 & 0 & 9\,774\,558 & 13\,430\,287 & yes \\ ToN-IoT & DDoS & 6\,436\,956 & 4\,141\,256 & 0 & 16\,545 & 4\,124\,711 & 274\,424 & yes \\ ToN-IoT & DoS & 7\,236\,417 & 203\,456 & 0 & 4 & 203\,452 & 350\,395 & yes \\ BoT-IoT & Benign & 31\,227\,838 & 51\,989 & 29 & 0 & 51\,960 & 14\,864\,652 & yes \\ BoT-IoT & DDoS & 3\,725\,689 & 7\,150\,882 & 0 & 0 & 7\,150\,882 & 488\,168 & yes \\ BoT-IoT & DoS & 188\,502 & 8\,034\,190 & 0 & 0 & 8\,034\,190 & 117\,981 & yes \\ BoT-IoT & Reconnaissance & 15\,383 & 1\,695\,132 & 0 & 120\,059 & 1\,575\,073 & 10\,744 & yes \\ CSE-CIC-2018 & Benign & 15\,618\,162 & 17\,392\,754 & 904\,743 & 0 & 16\,488\,011 & 2\,500\,316 & yes \\ \end{longtable}} \section{Leave-one-source-out: binary labels}\label{app:eval-p4b} Table~\ref{tab:app-eval-p4b} lists the labels that pass the thresholds for each held-out source, with the training rows (all other sources), the test rows (held-out source), the test rows that duplicate a training vector with the same label (``Same) or only with another label (``Other), the clean test rows, the distinct feature-set-B training vectors (``Dist.\ B) and whether the label is evaluable for feature set B. {\tiny \setlength{\tabcolsep}{2pt} \begin{longtable}{llrrrrrrc} \caption{Leave-one-source-out, binary labels: rows per evaluable label.}\label{tab:app-eval-p4b}\\ \toprule Held out & Label & Train & Test & Same & Other & Clean & Dist.\ B & Eval B \\ \midrule \endfirsthead \toprule Held out & Label & Train & Test & Same & Other & Clean & Dist.\ B & Eval B \\ \midrule \endhead \bottomrule \endfoot UNSW-NB15 & Benign & 29\,274\,486 & 2\,222\,930 & 1 & 0 & 2\,222\,929 & 13\,840\,904 & yes \\ UNSW-NB15 & Attack & 26\,641\,235 & 127\,679 & 0 & 0 & 127\,679 & 5\,346\,399 & yes \\ ToN-IoT & Benign & 17\,703\,512 & 15\,078\,514 & 5\,303\,956 & 331\,361 & 9\,443\,197 & 13\,430\,287 & yes \\ ToN-IoT & Attack & 17\,194\,358 & 10\,625\,609 & 75\,248 & 684\,485 & 9\,865\,876 & 1\,483\,923 & yes \\ BoT-IoT & Benign & 31\,227\,838 & 51\,989 & 29 & 1 & 51\,959 & 14\,864\,652 & yes \\ BoT-IoT & Attack & 11\,560\,867 & 16\,881\,819 & 114\,757 & 5\,315 & 16\,761\,747 & 4\,788\,532 & yes \\ CSE-CIC-2018 & Benign & 15\,618\,162 & 17\,392\,754 & 904\,743 & 143 & 16\,487\,868 & 2\,500\,316 & yes \\ CSE-CIC-2018 & Attack & 24\,871\,834 & 2\,094\,301 & 27\,914 & 9\,262 & 2\,057\,125 & 4\,685\,142 & yes \\ \end{longtable}} \section{Leave-one-source-out: families of taxonomy v1}\label{app:eval-p5} Table~\ref{tab:app-eval-p5} lists the labels that pass the thresholds for each held-out source, with the training rows (all other sources), the test rows (held-out source), the test rows that duplicate a training vector with the same label (``Same) or only with another label (``Other), the clean test rows, the distinct feature-set-B training vectors (``Dist.\ B) and whether the label is evaluable for feature set B. {\tiny \setlength{\tabcolsep}{2pt} \begin{longtable}{llrrrrrrc} \caption{Leave-one-source-out, families of taxonomy v1: rows per evaluable label.}\label{tab:app-eval-p5}\\ \toprule Held out & Label & Train & Test & Same & Other & Clean & Dist.\ B & Eval B \\ \midrule \endfirsthead \toprule Held out & Label & Train & Test & Same & Other & Clean & Dist.\ B & Eval B \\ \midrule \endhead \bottomrule \endfoot UNSW-NB15 & BENIGN & 29\,274\,486 & 2\,222\,930 & 1 & 0 & 2\,222\,929 & 13\,840\,904 & yes \\ UNSW-NB15 & COMMAND\_AND\_CONTROL & 191\,914 & 4\,658 & 0 & 0 & 4\,658 & 21\,013 & yes \\ UNSW-NB15 & DENIAL\_OF\_SERVICE & 18\,995\,598 & 39\,787 & 0 & 0 & 39\,787 & 1\,808\,176 & yes \\ UNSW-NB15 & EXPLOIT\_PUBLIC\_APP & 2\,896\,158 & 43\,970 & 0 & 0 & 43\,970 & 2\,650\,179 & yes \\ UNSW-NB15 & RECON\_DISCOVERY & 2\,748\,710 & 17\,074 & 0 & 0 & 17\,074 & 67\,906 & yes \\ ToN-IoT & BENIGN & 17\,703\,512 & 15\,078\,514 & 5\,303\,956 & 78\,379 & 9\,696\,179 & 13\,430\,287 & yes \\ ToN-IoT & COMMAND\_AND\_CONTROL & 104\,665 & 101\,692 & 0 & 2\,842 & 98\,850 & 6\,126 & yes \\ ToN-IoT & CREDENTIAL\_ACCESS & 260\,107 & 1\,600\,045 & 47 & 7\,655 & 1\,592\,343 & 85\,343 & yes \\ ToN-IoT & DENIAL\_OF\_SERVICE & 15\,122\,482 & 4\,344\,712 & 0 & 16\,549 & 4\,328\,163 & 1\,236\,048 & yes \\ ToN-IoT & EXPLOIT\_PUBLIC\_APP & 40\,426 & 3\,216\,212 & 0 & 75 & 3\,216\,137 & 37\,703 & yes \\ ToN-IoT & RECON\_DISCOVERY & 1\,541\,123 & 1\,358\,977 & 3\,001 & 728\,611 & 627\,365 & 36\,775 & yes \\ BoT-IoT & BENIGN & 31\,227\,838 & 51\,989 & 29 & 1 & 51\,959 & 14\,864\,652 & yes \\ BoT-IoT & DENIAL\_OF\_SERVICE & 5\,363\,300 & 15\,185\,072 & 0 & 0 & 15\,185\,072 & 1\,217\,492 & yes \\ BoT-IoT & RECON\_DISCOVERY & 1\,238\,353 & 1\,695\,132 & 114\,708 & 5\,361 & 1\,575\,063 & 52\,729 & yes \\ CSE-CIC-2018 & BENIGN & 15\,618\,162 & 17\,392\,754 & 904\,743 & 121 & 16\,487\,890 & 2\,500\,316 & yes \\ CSE-CIC-2018 & COMMAND\_AND\_CONTROL & 95\,683 & 111\,460 & 0 & 0 & 111\,460 & 21\,931 & yes \\ CSE-CIC-2018 & CREDENTIAL\_ACCESS & 1\,439\,560 & 289\,215 & 1 & 16 & 289\,198 & 638\,780 & yes \\ CSE-CIC-2018 & DENIAL\_OF\_SERVICE & 17\,612\,589 & 1\,576\,921 & 0 & 8\,325 & 1\,568\,596 & 1\,250\,034 & yes \\ CSE-CIC-2018 & EXPLOIT\_PUBLIC\_APP & 2\,934\,938 & 900 & 0 & 0 & 900 & 2\,687\,648 & yes \\ \end{longtable}} \section{Leave-one-source-out: families of taxonomy v0}\label{app:eval-p5v0} Table~\ref{tab:app-eval-p5v0} lists the labels that pass the thresholds for each held-out source, with the training rows (all other sources), the test rows (held-out source), the test rows that duplicate a training vector with the same label (``Same) or only with another label (``Other), the clean test rows, the distinct feature-set-B training vectors (``Dist.\ B) and whether the label is evaluable for feature set B. {\tiny \setlength{\tabcolsep}{2pt} \begin{longtable}{llrrrrrrc} \caption{Leave-one-source-out, families of taxonomy v0: rows per evaluable label.}\label{tab:app-eval-p5v0}\\ \toprule Held out & Label & Train & Test & Same & Other & Clean & Dist.\ B & Eval B \\ \midrule \endfirsthead \toprule Held out & Label & Train & Test & Same & Other & Clean & Dist.\ B & Eval B \\ \midrule \endhead \bottomrule \endfoot UNSW-NB15 & BENIGN & 29\,274\,486 & 2\,222\,930 & 1 & 0 & 2\,222\,929 & 13\,840\,904 & yes \\ UNSW-NB15 & FLOOD\_DOS & 18\,995\,598 & 5\,971 & 0 & 0 & 5\,971 & 1\,808\,176 & yes \\ UNSW-NB15 & MALWARE\_C2 & 195\,482 & 4\,816 & 0 & 0 & 4\,816 & 23\,574 & yes \\ UNSW-NB15 & RECON\_SCAN & 2\,748\,710 & 18\,300 & 0 & 0 & 18\,300 & 67\,906 & yes \\ ToN-IoT & BENIGN & 17\,703\,512 & 15\,078\,514 & 5\,303\,956 & 78\,379 & 9\,696\,179 & 13\,430\,287 & yes \\ ToN-IoT & BRUTE\_FORCE & 260\,107 & 1\,594\,777 & 47 & 7\,655 & 1\,587\,075 & 85\,343 & yes \\ ToN-IoT & FLOOD\_DOS & 15\,092\,107 & 4\,344\,712 & 0 & 16\,549 & 4\,328\,163 & 1\,211\,793 & yes \\ ToN-IoT & MALWARE\_C2 & 104\,807 & 105\,663 & 0 & 2\,883 & 102\,780 & 6\,251 & yes \\ ToN-IoT & RECON\_SCAN & 1\,542\,235 & 1\,358\,977 & 3\,001 & 728\,611 & 627\,365 & 37\,887 & yes \\ ToN-IoT & WEB\_INJECTION & 823 & 3\,216\,212 & 0 & 75 & 3\,216\,137 & 117 & yes \\ BoT-IoT & BENIGN & 31\,227\,838 & 51\,989 & 29 & 1 & 51\,959 & 14\,864\,652 & yes \\ BoT-IoT & FLOOD\_DOS & 5\,332\,925 & 15\,185\,072 & 0 & 0 & 15\,185\,072 & 1\,193\,237 & yes \\ BoT-IoT & INFIL\_EXFIL & 104\,152 & 1\,615 & 0 & 3 & 1\,612 & 74\,375 & yes \\ BoT-IoT & RECON\_SCAN & 1\,239\,465 & 1\,695\,132 & 114\,708 & 5\,361 & 1\,575\,063 & 53\,841 & yes \\ CSE-CIC-2018 & BENIGN & 15\,618\,162 & 17\,392\,754 & 904\,743 & 121 & 16\,487\,890 & 2\,500\,316 & yes \\ CSE-CIC-2018 & BRUTE\_FORCE & 1\,434\,832 & 289\,215 & 1 & 16 & 289\,198 & 634\,270 & yes \\ CSE-CIC-2018 & FLOOD\_DOS & 17\,582\,214 & 1\,576\,921 & 0 & 8\,325 & 1\,568\,596 & 1\,225\,779 & yes \\ CSE-CIC-2018 & INFIL\_EXFIL & 1\,468 & 115\,805 & 0 & 28\,834 & 86\,971 & 411 & yes \\ CSE-CIC-2018 & MALWARE\_C2 & 99\,393 & 111\,460 & 0 & 0 & 111\,460 & 24\,617 & yes \\ CSE-CIC-2018 & WEB\_INJECTION & 2\,895\,335 & 900 & 0 & 0 & 900 & 2\,650\,062 & yes \\ \end{longtable}}