Abnormal sample detection based on robust Mahalanobis distance estimation in adversarial machine learning

Abstract

This paper addresses the problem of abnormal sample detection in deep learning-based computer vision, focusing on two types of abnormal samples: outlier samples and adversarial samples. The presence of these abnormal samples can significantly degrade the performance and robustness of deep learning models, posing security risks in critical areas. To address this, we propose a method that combines robust Mahalanobis distance (RMD) estimation with a pretrained convolutional neural networks (CNNs) model. The RMD estimation involves using minimum covariance matrix determinant (MCD), $T$-type, and $S$ estimators. Furthermore, we theoretically analyze the breakdown point and influence function of the $T$-type estimator. To evaluate the effectiveness and robustness of our method, we utilize public datasets, CNN models, and adversarial sample generation algorithms commonly employed in the field. The experimental results demonstrate the effectiveness of our algorithm in detecting abnormal samples.