Browse Teach Curate

Attack-as-Defense Backdoor Injection Against Model Extraction in Deep Learning

A defense paradigm for protecting black-box classifiers from model extraction: instead of only blocking or watermarking, the victim exposes a honeypot output layer whose poisoned probability vectors implant a backdoor into any substitute model an attacker trains, enabling ownership verification and a trigger-driven reverse attack.

Data Text

HoneypotNet: Backdoor Attacks Against Model Extraction Yixu Wang1, 2*, Tianle Gu2, 3, Yan Teng2†

Model extraction attacks steal a black-box image classifier by querying its API and training a substitute model on the returned prediction vectors. This paper introduces the 'attack as defense' parad…