SparSEEty: Extracting Tokens from Sparsity-Exploiting LLM Serving Systems via Deterministic Side Channels

2026-08-03

AI Summary

This work demonstrates, for the first time, an end-to-end deterministic side-channel attack capable of extracting input and output tokens from sparsity-optimized large language models within Intel TDX confidential virtual machines. Exploiting input-dependent weight access patterns inherent to sparse inference, the attack constructs a low-overhead neuron activation oracle by fusing multiple side-channel signals exposed by TDX. To overcome challenges such as high monitoring overhead and the difficulty of inverting binary activations, the authors design an efficient activation trace inversion algorithm. Experimental results across diverse models and datasets show that the method achieves highly accurate token reconstruction—exceeding BLEU scores of 0.95—while incurring only 3.7%–7.2% monitoring overhead.

Abstract

Modern large language models (LLMs) exhibit activation sparsity, wherein only a subset of their neurons is activated for given input tokens. Researchers have leveraged this property to optimize LLM serving systems by omitting weight accesses and computations pertaining to inactive neurons. Unfortunately, however, such optimizations create input-dependent weight accesses, which can be leaked over side channels. We present SparSEEty, a new token extraction attack that exploits input-dependent neuron weight accesses introduced by sparsity-exploiting LLM serving systems. SparSEEty first constructs a neuron-activation oracle using neuron weight access side channels during LLM inference, and then inverts the activation traces to reconstruct the input tokens, forming an end-to-end token extraction attack. We instantiate SparSEEty against an LLM serving system protected inside an Intel TDX confidential virtual machine (CVM), addressing three key challenges: (i) constructing a neuron-activation oracle using a combination of side channels exposed by CVMs, (ii) reducing inference-time overheads of neuron activation monitoring for covertness, and (iii) accurately inverting partial binary activation traces back to tokens. Our evaluation shows that SparSEEty can reconstruct both prompt and response tokens with consistently high BLEU scores (>0.95) across various models and datasets, while incurring monitoring overheads of 3.7% to 7.2%.

Problem

• side channels
• activation sparsity
• token extraction
• LLM serving
• confidential computing

Innovation

• activation sparsity
• side-channel attack
• token extraction
• confidential computing
• LLM serving