Multi-Target Generative Adversarial Attacks Based on Dual-Information Alignment

Deep neural networks, as a frontier technology in the field of artificial intelligence, have achieved remarkable success across various domains. However, recent studies have revealed their susceptibility to adversarial attacks, especially targeted attacks that can precisely control the output of unknown models, posing significant threats to data privacy and system security. Generative attack methods, known for efficiently crafting adversarial examples, have increasingly been employed in targeted attack research. Nevertheless, existing generative attack methods primarily focus on adversarial examples for single target classes, leading to inefficiencies, limited flexibility, and poor scalability in multi-target tasks. To address these challenges, this paper proposes a Multi-Target Generative Attack based on Dual-Information (MTGA-DI). This method leverages a conditional generative model to simultaneously utilize the semantic and visual information of target classes, enabling effective multi-target attack capabilities while significantly improving the transferability and robustness of adversarial examples. Experimental results demonstrate that MTGA-DI outperforms previous multi-target attack methods on both standard and robustly trained models, and achieves superior performance against models employing input preprocessing defenses.