Improving out-of-distribution image detection via second eigenvector perturbation

Abstract

Although deep neural network has high performance on image classification, it suffers from the lack of ability to distinguish between normal inputs and abnormal inputs, e.g. adversarial examples and out-of-distribution samples. Since deep neural network with softmax classifier is overconfident, detecting out-of-distribution samples only with softmax prediction seems to be insufficient. ODIN is the baseline approach to detect out-of-distribution samples by adding small perturbation which increases the softmax score. This preprocessing technique is widely adopted on state-of-the-art out-of-distribution detection methods. Despite the efficiency of ODIN, its linear approximation restricts theoretical approach and curvature analysis. In this work, we propose additional input preprocessing method named Second Eigenvector Perturbation which observes local curvature by Hessian matrix. Our method improves out-of-distribution detection performance compared with ODIN. Furthermore, we visualize data landscapes to give intuition of the effect of input preprocessing and demonstrate that out-of-distribution with irregular landscape experiences input preprocessing differently.