Llm-ga: A gradient-based multi-label adversarial attack by large language models

Abstract

Deep neural networks (DNNs) are highly sensitive to small, meticulously crafted perturbations, which have been utilized in adversarial attacks, threatening the reliability of DNNs in practical applications. Current adversarial attack methods rely heavily on expert design, requiring significant researcher effort. In this paper, we introduce LLM-GA, a Large Language Model-based Gradient Attack method, specifically designed for generating adversarial attacks against multi-label classification models. Based on the prompts and algorithm templates provided by attackers, LLM-GA can automatically generate ideas for attack algorithms and produce corresponding code implementations. This significantly improves the efficiency of designing adversarial attack algorithms. To generate more effective attack algorithms, LLM-GA leverages evolutionary algorithms to iteratively refine its ideas. Experimental results demonstrate that LLM-GA outperforms three existing gradient-based adversarial attacks in terms of both attack success rate and perturbation size. The code for this paper is available https://github.com/liuyujiang123/LLM-GA.