Sustainable Adversarial Attacks Based on Semantic Enhancement in Class-Incremental Learning

Key Points

• The study aims to improve the robustness of adversarial samples in class-incremental learning by using semantic enhancement techniques.
• Proposed a targeted adversarial sample generation method based on semantic reinforcement.
• Utilized diffusion models for candidate sample generation and initial model prediction entropy for semantic filtering.
• Designed semantic constraint functions to ensure adversarial samples are close to target classes and away from non-target classes.
• Achieved an average attack success rate improvement of 39.38% compared to baseline methods when category scale increased tenfold.