Attacking logic rather than outputs: a concept-attribute attack that eliminates neural network shortcuts

The paper proposes a concept-attribute attack against neural networks, extending adversarial attacks from falsifying output classes to falsifying logical constraints on human-interpretable concepts, for example that red and octagonal should imply a stop sign. The authors implement the method on trained models using concept-based explainable AI (C-XAI) techniques, and formally prove that such attacks are expected to have a smaller search space than simple class falsification. In experiments across multiple models and datasets covering vision tasks and a security-related code processing task, retraining based on the new attack better eliminates semantic bias while maintaining the robustness gains of classic adversarial retraining.