<ul><li>This paper introduces a new side-channel in large language models (LLMs) that allows an adversary to extract sensitive information about inference inputs.</li><li>The side-channel is based on the number of output tokens in the LLM response.</li><li>The paper demonstrates attacks utilizing this side-channel in machine translation tasks and text classification tasks.</li><li>Proposed mitigations against the output token count side-channel are also discussed.</li></ul>

Time Will Tell: Timing Side Channels via Output Token Count in Large Language Models

Discover more