notesum.ai

Published at November 26

PEFTGuard: Detecting Backdoor Attacks Against Parameter-Efficient Fine-Tuning

cs.CR

Released Date: November 26, 2024

Authors: Zhen Sun¹, Tianshuo Cong², Yule Liu¹, Chenhao Lin³, Xinlei He¹, Rongmao Chen⁴, Xingshuo Han⁵, Xinyi Huang⁶

Aff.: ¹The Hong Kong University of Science and Technology (Guangzhou); ²Tsinghua University; ³Xi'an Jiaotong University; ⁴National University of Defense Technology; ⁵Nanyang Technological University; ⁶Jinan University

Arxiv: http://arxiv.org/abs/2411.17453v1

Method	SC		IF
Method	Detection Acc	AUC	Detection Acc	AUC
T-Miner (USENIX’21)	$50\%$	$0.500$	-	-
AttenTD (NAACL’22)	$50\%$	$0.606$	-	-
PICCOLO (S&P’22)	$76\%$	$0.890$	-	-
MNTD (S&P’21)	$88\%$	$0.937$	$51\%$	$0.510$
$\mathsf{PEFTGuard}$ (Ours)	99%	1.000	100%	1.000