IATA: Instance-driven Advancing Targeted Attacks with Transferable Pattern Embedding

Generative attack methods have attracted increasing attention in recent years due to their potential to efficiently deceive black-box models. Among the various attacks, the targeted attacks focus on misleading victim models to produce adversary-desired predictions, which are more challenging and threatening than the untargeted attacks. However, the current methods still exhibit compromised generalization in low-data regimes due to insufficient disentanglement of the instance-level discriminative features from dataset-wide adversarial pattern distributions. To this end, we propose an instance-driven advancing targeted attacks (IATA) framework with transferable pattern embedding. IATA achieves high transferability by integrating dual-branch pattern injection and local patch guided adversarial attacks. Specifically, the instance-driven adversarial generator and the prototype discriminator are designed to fuse instance-level embeddings from a specific target and the prototype-level features of the target class, respectively. The local patch tuning strategy is designed to enhance the dual-substitute model discrepancy attack by stimulating more texture perturbations. Extensive experiments demonstrate that IATA achieves stateof-the-art performance in the black-box setting and significantly outperforms existing targeted attack methods by a margin of 14.98 %, as measured by the targeted transfer attack success rate when Inception-v3 is used as the substitute model.