On the receptive field misalignment in CAM-based visual explanations

摘要

Visual explanations aim at providing an understanding of the inner behavior of convolutional neural net-works. Naturally, it is necessary to explore whether these methods themselves are reasonable and reli-able. In this paper, we focus on Class Activation Mapping (CAM), a type of attractive explanations that has been widely applied to model diagnosis and weakly supervised tasks. Our contribution is two-fold. First, we identify an important but neglected issue that affects the reliability of CAM results: there is a misalignment between the effective receptive field and the implicit receptive field, where the former is determined by the model and the input, and the latter is determined by the upsampling function in CAM. Occlusion experiments are designed to empirically testify to its existence. Second, based on this finding, an adversarial marginal attack is proposed to fool the CAM-based method and the CNN model simulta-neously. Experimental results demonstrate that the provided saliency map can be completely changed to another shape by only perturbing the area with 1-pixel width. The prototype code of the method is now available at https://github.com/xpf/CAM- Adversarial- Marginal-Attack . (c) 2021 Elsevier B.V. All rights reserved.