A Transferable Targeted Adversarial Attack Method Based on Feature Interpretation and Dynamic Optimization

Deep neural networks (DNNs) are widely applied in critical fields but remain vulnerable to adversarial attacks. Among them, transferable targeted adversarial attacks in black-box scenarios are particularly dangerous, as they can mislead models into predicting specified target labels. Existing methods improve targeted transferability through optimization refinement, input diversity, source-model enhancement, or feature-space perturbation. However, they generally do not exploit interpretation information from a white-box source model (WSM) as iterative feedback to refine global perturbations toward discriminative local regions. Thus, redundant and source-model-specific components may remain in less informative regions, limiting cross-model transferability. To address these limitations, we propose a transferable targeted adversarial attack method based on feature interpretation and dynamic optimization (TAFID). TAFID introduces an interpretation-feedback optimization mechanism within an iterative transfer-based targeted attack framework. It uses global-local feature alignment to maximize the target-label output and the cosine similarity between global and cropped local features, promoting stable target-oriented intermediate representations. Subsequently, it applies a dynamically updated perturbation mask to reweight the global perturbation according to the changing focus area of the WSM. Experimental results demonstrate that when attacking DNNs, third-party deep learning systems and defense methods, TAFID improves the average targeted attack success rate (TASR) on black-box target models (BTMs) by 48.6%, 30.6%, 30.7%, 42.8%, 48.3%, and 36.4% compared with reverse adversarial perturbation (RAP), object-based diverse input (ODI), sharpness-aware self-distillation and weight scaling (SASD-WS), clean feature mixup (CFM), SupHigh+ft and feature tuning mixup (FTM), respectively. Moreover, the generated adversarial examples maintain reasonable similarity to the original examples, indicating a favorable balance between attack effectiveness and example authenticity. Our code is available at https://github.com/zyt2020/TAFID.