Secure Speculative Decoding for Large Language Models
Lossy speculative decoding can speed LLM inference while sharply weakening resistance to jailbreaks and prompt injection.
The paper says those attack success rates rise much faster than ordinary utility falls across tested lossy speculative decoding methods. Its explanation is that much of the security loss comes from early tokens proposed by the smaller draft model. The authors propose SecureSD, which applies stricter verification to draft tokens at early decoding positions. They report that it improves security while preserving efficiency and utility against existing methods. ArXiv · AI/CL/LG's note
The paper says those attack success rates rise much faster than ordinary utility falls across tested lossy speculative decoding methods. Its explanation is that much of the security loss comes from early tokens proposed by the smaller draft model. The authors propose SecureSD, which applies stricter verification to draft tokens at early decoding positions. They report that it improves security while preserving efficiency and utility against existing methods. ArXiv · AI/CL/LG's note
score 5