Adversarial attacks pose serious challenges for deep neural network (DNN)-based analysis of various input signals. In the case of 3D point clouds, methods have been developed to identify points that play a key role in network decision, and these become crucial in generating existing adversarial attacks. For example, a saliency map approach is a popular method for identifying adversarial drop points, whose removal would significantly impact the network decision. Generally, methods for identifying adversarial points rely on the access to the DNN model itself to determine which points are critically important for the model's decision. This paper aims to provide a novel viewpoint on this problem, where adversarial points can be predicted without access to the target DNN model, which is referred to as a ``no-box'' attack. To this end, we define 14 point cloud features and use multiple linear regression to examine whether these features can be used for adversarial point prediction, and which combination of features is best suited for this purpose. Experiments show that a suitable combination of features is able to predict adversarial points of four different networks -- PointNet, PointNet++, DGCNN, and PointConv -- significantly better than a random guess and comparable to white-box attacks. Additionally, we show that no-box attack is transferable to unseen models. The results also provide further insight into DNNs for point cloud classification, by showing which features play key roles in their decision-making process.
翻译:暂无翻译