SparSEEty: Extracting Tokens from Sparsity-Exploiting LLM Serving Systems
SparSEEty addresses token extraction attacks on LLM serving systems, achieving high reconstruction accuracy with minimal overhead.
Modern large language models (LLMs) exhibit activation sparsity, allowing for optimizations that omit inactive neurons. However, these optimizations lead to input-dependent weight accesses, which can be exploited through side channels. SparSEEty is a novel token extraction attack that constructs a neuron-activation oracle using these side channels during LLM inference, enabling the reconstruction of input tokens. Evaluations show that SparSEEty achieves high BLEU scores (>0.95) in reconstructing both prompt and response tokens across various models and datasets, with monitoring overheads ranging from 3.7% to 7.2%.