C-FA: Enhancing Transferability Through Contrastive Learning and Feature Attention in Point Cloud Attacks

3D classification networks are vulnerable to adversarial examples, which can cross decision boundaries and cause the model to misclassify. However, the adversarial examples generated by the current method do not transferred well across the decision boundaries of various models, resulting in a lower success rate of transfer to a black-box model and poor attack performance against the target model. In this paper, we propose a new method, called C-FA, which revisits the problem of crossing model decision boundaries, generating adversarial examples with improved transferability. We use a contrastive learning strategy, by increasing the difference between the adversarial examples and the benign examples in the feature space, prompting the adversarial examples to cross the decision boundary, leading to misclassification by the model. In addition, considering that perturbing the salient features focused on by the model can effectively enhance the transferability of the generated adversarial examples, we introduce a feature attention mechanism that identifies and weights features that have a more significant impact on decision-making by analyzing the feature gradient. The effectiveness of C-FA and its competitiveness against existing state-of the-art techniques are confirmed through extensive experiments on six representative networks including PointNet, PointNet++, DGCNN, PointConv, PCT, and Mamba3D.