notesum.ai

Published at November 5

Membership Inference Attacks against Large Vision-Language Models

cs.CV

cs.AI

cs.CL

cs.CR

cs.LG

Released Date: November 5, 2024

Authors: Zhan Li¹, Yongtao Wu¹, Yihang Chen¹, Francesco Tonin¹, Elias Abad Rocamora¹, Volkan Cevher¹

Aff.: ¹LIONS, EPFL

Arxiv: http://arxiv.org/abs/2411.02902v1

Refer to caption

Metric		VLLM Tuning		LLM Pre-Training
Metric		32	64	32	64	128	256
Perplexity^∗		0.779	0.988	0.542	0.505	0.553	0.582
Perplexity/zlib^∗		0.609	0.986	0.56	0.537	0.581	0.603
Perplexity/lowercase^∗		0.962	0.977	0.493	0.518	0.503	0.583
Min_0% Prob^∗		0.522	0.522	0.455	0.451	0.425	0.448
Min_10% Prob^∗		0.461	0.883	0.468	0.487	0.526	0.534
Min_20% Prob^∗		0.603	0.980	0.505	0.498	0.549	0.562
Max_Prob_Gap		0.461	0.545	0.574	0.544	0.565	0.629
ModRényi^∗	$\alpha=0.5$	0.809	0.979	0.557	0.500	0.536	0.567
	$\alpha=1$	0.808	0.993	0.544	0.503	0.546	0.567
	$\alpha=2$	0.779	0.963	0.559	0.497	0.529	0.560
Rényi ( $\alpha=0.5$ )	Max_0%	0.506	0.514	0.541	0.515	0.489	0.571
	Max_10%	0.458	0.776	0.518	0.525	0.606	0.65
	Max_100%	0.564	0.835	0.555	0.531	0.6	0.631
Rényi ( $\alpha=1$ )	Max_0%	0.552	0.579	0.566	0.571	0.603	0.668
	Max_10%	0.566	0.809	0.553	0.541	0.623	0.65
	Max_100%	0.554	0.750	0.544	0.523	0.588	0.621
Rényi ( $\alpha=2$ )	Max_0%	0.589	0.625	0.594	0.606	0.659	0.657
	Max_10%	0.607	0.787	0.583	0.556	0.629	0.663
	Max_100%	0.553	0.709	0.592	0.576	0.568	0.649
Rényi ( $\alpha=\infty$ )	Max_0%	0.600	0.638	0.607	0.615	0.688	0.669
	Max_10%	0.618	0.763	0.586	0.548	0.627	0.667
	Max_100%	0.557	0.694	0.546	0.527	0.584	0.634