Side-Channel Attacks Targeting Language Models Detailed in Three Studies

Three studies detail side-channel attacks targeting language models (LLMs). The first, "Remote Timing Attacks on Efficient Language Model Inference," exploits response time variations in optimized systems (such as speculative sampling) to deduce the content of messages through encrypted network traffic. It achieves 90% accuracy in identifying conversation themes (e.g., medical vs. coding) and retrieves personal data (PII) from open-source models.