Offensive AI Security› Extraction Attacks›Stealing the Last Layer: Recovering a Production Model's Final Projection from Logits

In this series

Extraction Attacks

1The Query-Access Threat Model: What a Black-Box Attacker Can and Cannot LearnResearch
2Stealing the Last Layer: Recovering a Production Model's Final Projection from LogitsResearch
3Training-Data Extraction and Memorization: Eidetic Memorization and the Extraction RateResearch
4Membership Inference: Shadow Models, Loss Thresholds, and Calibrated AUC as a Privacy MetricResearch
5Defenses and Their Costs: Truncation, Noise, Rate Limits, and Differential-Privacy AccountingResearch

← All research

HostExpert

Murali Chillakuru

Press play to begin the walkthrough.

0:00

15:21

Extraction Attacks · 2 / 5

Stealing the Last Layer: Recovering a Production Model's Final Projection from Logits

Back