Smoothing Geometry for Robust Attribution of Deep Neural Networks

The researchers propose a method to mitigate adversarial attacks on feature attributions by introducing a Lipschitz continuity condition on model gradients, making gradient-based attributions more robust. Such attacks can cause nearby inputs to produce very different explanations, harming safety and trustworthiness in high-stakes applications. The paper presents a low-cost regularization method and a randomized smoothing technique that requires no retraining; experiments on multiple types of image models show that both consistently improve attribution robustness, and on real large-scale models they confirm the role of smoothing geometry in such attacks.