SpecGuard: Inference-Time Backdoor Detection For Free
SpecGuard uses speculative decoding’s own token-acceptance signal to flag triggered backdoor behavior without extra model runs.
The paper says a clean draft model and a compromised target model diverge when a hidden trigger pushes the target toward attacker-controlled output. That divergence changes the draft-token acceptance rate, giving the detector its runtime signal. The authors report detection across multiple backdoor types and model families, including cases missed by input-level filters. ArXiv · AI/CL/LG's note
The paper says a clean draft model and a compromised target model diverge when a hidden trigger pushes the target toward attacker-controlled output. That divergence changes the draft-token acceptance rate, giving the detector its runtime signal. The authors report detection across multiple backdoor types and model families, including cases missed by input-level filters. ArXiv · AI/CL/LG's note
score 5