IoTProber's picture
Trim the artifact-gap note from the drift README
fc1a10c verified
|
Raw History Blame Contribute Delete
3.54 kB

SVM_validation — Drift Evaluation Ground Truth & Correctness Criteria

This directory holds the drift-detection evaluation referenced by the paper's drift-correctness table. A reviewer noted that some IPs listed in IoTProber/incorrect_ips.csv also appear in Ground/drift_ips.csv, which looks contradictory. It is not — membership and correctness answer two different questions at two different stages. This file documents the exact semantics.

Directory layout

Path What it is
Ground/class_drift_scores.csv Class-level ground truth. One row per (device_type, vendor) class: drift_score, n_old, n_new, drifted (True/False). 182 classes total; 65 drifted=True, 117 drifted=False.
Ground/drift_ips.csv IP-level expansion of the drifted classes: all 1,357 validation IPs whose (device_type, vendor) class has drifted=True. Verified: its 65 distinct classes are exactly the 65 drifted=True classes; no drifted=False class contributes an IP.
IoTProber/detected_ips.csv The 814 IPs flagged as drifted by IoTProber's drift detector in this evaluation.
IoTProber/correct_ips.csv 773 of the 814 flagged IPs — see correctness criterion below.
IoTProber/incorrect_ips.csv The other 41 flagged IPs.
performance/IoTProber/predict/*_val_pred.csv A separate classification benchmark population (7,239 IPs, ip,vendor,label,label_pred,correct). Disjoint from the 814 drift IPs — do not join against it.

The two-stage semantics

Stage 1 — drift detection (membership). The detection evaluation runs over the population of IPs from drifted=True classes (drift_ips.csv). Therefore:

  • Every flagged IP belongs to a ground-truth drifted class by construction: detection precision is 100% by design, not an empirical result.
  • Detection recall = 814 / 1,357 = 60.0%; the 543 unflagged ground-truth IPs are the misses.

Stage 2 — post-drift re-identification (correctness). For each flagged IP, the pipeline re-identifies the device; the IP is

  • correct — the post-drift re-identification matches the ground-truth (type, vendor) label: 773 / 814 = 95.0% (the paper's reported percentage);
  • incorrect — the re-identification mismatches (41 IPs, concentrated in 7 classes; PRINTER×25, BUILDING_AUTOMATION×8, CAMERA×3, NVR×3, MEDICAL×1, ROUTER×1).

Consequently, an IP can be in drift_ips.csv and in incorrect_ips.csv simultaneously without contradiction: the first says "it truly drifted (Stage-1 membership)", the second says "the detector flagged it correctly but its post-drift re-identification was wrong (Stage-2)".

Column semantics of correct/incorrect/detected CSVs

All four columns — ip, vendor, device_type, class_drift_score — are ground-truth attributes of the listed IPs (joinable to Ground/drift_ips.csv; values verified identical). They are not detector outputs. The correct/incorrect split itself is the only judgment column (implied by the file membership).

Quick reproduction

import pandas as pd
det = pd.read_csv("IoTProber/detected_ips.csv")
cor = pd.read_csv("IoTProber/correct_ips.csv")
inc = pd.read_csv("IoTProber/incorrect_ips.csv")
gt  = pd.read_csv("Ground/drift_ips.csv")
assert len(cor) + len(inc) == len(det) == 814
assert set(det["ip"]) <= set(gt["ip"])          # membership: all flagged are truth
print("recall      =", len(det) / len(gt))       # 0.600
print("correctness =", len(cor) / len(det))      # 0.950