notesum.ai

Published at November 12

Zer0-Jack: A Memory-efficient Gradient-based Jailbreaking Method for Black-box Multi-modal Large Language Models

cs.LG

cs.AI

Released Date: November 12, 2024

Authors: Tiejin Chen¹, Kaishen Wang¹, Hua Wei¹

Aff.: ¹School of Computing and Augmented Intelligence, Arizona State University, USA

Arxiv: http://arxiv.org/abs/2411.07559v1

Refer to caption

Model	P-Text	GCG	AutoDAN	PAIR	G-Image	P-Image	A-Image	WB	\ours
MiniGPT-4	11%	13%	16%	14%	10%	11%	13%	93%	95%
LLaVA1.5	0	0	8%	5%	0	1%	0	91%	90%
INF-MLLM1	0	1%	22%	7%	0	1%	1%	86%	88%
MiniGPT-4 (70B)	14%	-	-	17%	12%	13%	-	-	92%