notesum.ai

Published at November 22

Who Can Withstand Chat-Audio Attacks? An Evaluation Benchmark for Large Language Models

cs.SD

cs.AI

eess.AS

Released Date: November 22, 2024

Authors: Wanqi Yang¹, Yanda Li¹, Meng Fang², Yunchao Wei³, Tianyi Zhou⁴, Ling Chen¹

Aff.: ¹University of Technology Sydney; ²University of Liverpool; ³Beijing Jiaotong University; ⁴Agency for Science, Technology and Research (A*STAR), Singapore

Arxiv: http://arxiv.org/abs/2411.14842v1

$Refer to caption$

Audio Attack		MELD	TVQA	Common Voice
No Attack		120	120	120
Content Attack		120	120	120
Emotion Attack	Opp-Emo Tone	120	-	-
Emotion Attack	Opp-Emo Music	120	-	-
Explicit Noise	Natural Noise	40	40	40
	Industrial Noise	40	40	40
	Human Noise	40	40	40
Implicit Noise	Infrasound	60	60	60
Implicit Noise	Ultrasound	60	60	60
Total		1,680