notesum.ai

Published at November 27

InputSnatch: Stealing Input in LLM Services via Timing Side-Channel Attacks

cs.CR

Released Date: November 27, 2024

Authors: Xinyao Zheng¹, Husheng Han¹, Shangyi Shi¹, Qiyan Fang², Zidong Du¹, Qi Guo¹, Xing Hu¹

Aff.: ¹SKLP, Institute of Computing Technology, Chinese Academy of Sciences; ²University of Science and Technology of China

Arxiv: http://arxiv.org/abs/2411.18191v1

Refer to caption

Ideal
Methods	$ASR_{disease}$	$ASR_{symptoms}$	$ASR_{all}$	Attempts	Tokens	Time
Baseline	60.00%	23.50%	22.00%	$6868\pm 6855$	$283799\pm 273969$	$1751\pm 1747$
GaussianNB	59.00%	27.50%	26.50%	$2904\pm 2830$	$108650\pm 94450$	$730\pm 622$
Prob_vocabulary	67.50%	53.75%	49.00%	$502720\pm 2426519$	$11507996\pm 56966635$	$77918\pm 376100$
Finetuned LLM	54.00%	18.00%	10.00%	$659920\pm 6817828$	$6599206\pm 6817828$	$101500\pm 101001$
All
Methods	$ASR_{disease}$	$ASR_{symptoms}$	$ASR_{all}$	Attempts	Tokens	Time
Baseline	45.50%	9.00%	8.00%	$1350\pm 1173$	$55926\pm 48139$	$232\pm 166$
GaussianNB	54.00%	13.50%	12.50%	$1086\pm 898$	$40207\pm 33421$	$276\pm 228$
Prob_vocabulary	62.00%	12.50%	9.50%	$2247\pm 983$	$57526\pm 26768$	$348\pm 152$
Finetuned LLM	00.00%	00.00%	00.00%	$0\pm 0$	$0\pm 0$	$0\pm 0$