notesum.ai

Published at November 29

VLSBench: Unveiling Visual Leakage in Multimodal Safety

cs.CR

cs.AI

cs.CL

cs.CV

Released Date: November 29, 2024

Authors: Xuhao Hu¹, Dongrui Liu¹, Hao Li², Xuanjing Huang³, Jing Shao¹

Aff.: ¹Shanghai Artificial Intelligence Laboratory; ²Shanghai Artificial Intelligence Laboratory, Beihang University; ³Fudan University

Arxiv: http://arxiv.org/pdf/2411.19939v1

Refer to caption

Baselines	Training Method	Learning Rate	Epochs	Max Length
LLaVA-v1.5-7B + Textual SFT	full-parameter	$2e-5$	$2$	$2048$
LLaVA-v1.5-7B + Textual Unlearning	full-parameter	$2e-5$	$3$	$2048$
LLaVA-v1.5-13B + Textual SFT	full-parameter	$2e-5$	$2$	$2048$
Qwen2-VL-7B + Textual SFT	full-parameter	$1e-5$	$2$	$1024$
Qwen2-VL-7B + Multimodal SFT	full-parameter	$1e-5$	$3$	$1024$