Dataset Poisoning Detector

Spectral signature + influence function detectors · Source repo ↗
Committed synthetic label-flip benchmark data. In results/spectral_benchmark.json, spectral detection averages F1=0.6142 versus 0.1531 for the feature-space ensemble across 5%, 10%, and 20% contamination. These are synthetic benchmark F1 values, not ROC-AUC or production efficacy.
Run live detection eval
Live AUC: —
0.6142
Avg spectral F1
5/10/20% label-flip contamination
0.1531
Avg ensemble F1
Z-score + IQR + Isolation Forest
3
Contamination rates
5% · 10% · 20%
2,000
Synthetic samples
100 features · seed 2018
Label-flip F1 by contamination rate
Committed benchmark comparison
Spectral · 5%
0.5500
Spectral · 10%
0.5950
Spectral · 20%
0.6975
Ensemble average
0.1531