Extreme Values are Accurate and Robust in Deep Networks

Recent evidence shows that convolutional neural networks (CNNs) are biased towards textures so that CNNs are non-robust to adversarial perturbations over textures, while traditional robust visual features like SIFT (scale-invariant feature transforms) are designed to be robust across a substantial range of affine distortion, addition of noise, etc with the mimic of human perception nature. This paper aims to leverage good properties of SIFT to renovate CNN architectures towards better accuracy and robustness. We borrow the scale-space extreme value idea from SIFT, and propose EVPNet (extreme value preserving network) which contains three novel components to model the extreme values: (1) parametric differences of Gaussian (DoG) to extract extrema, (2) truncated ReLU to suppress non-stable extrema and (3) projected normalization layer (PNL) to mimic PCA-SIFT like feature normalization. Experiments demonstrate that EVPNets can achieve similar or better accuracy than conventional CNNs, while achieving much better robustness on a set of adversarial attacks (FGSM,PGD,etc) even without adversarial training.