Robust Determinantal Generative Classifier for Noisy Labels and Adversarial Attacks

Large-scale datasets may contain significant proportions of noisy (incorrect) class labels, and it is well-known that modern deep neural networks poorly generalize from such noisy training datasets. In this paper, we propose a novel inference method, Deep Determinantal Generative Classifier (DDGC), which can obtain a more robust decision boundary under any softmax neural classifier pre-trained on noisy datasets. Our main idea is inducing a generative classifier on top of hidden feature spaces of the discriminative deep model. By estimating the parameters of generative classifier using the minimum covariance determinant estimator, we significantly improve the classification accuracy, with neither re-training of the deep model nor changing its architectures. In particular, we show that DDGC not only generalizes well from noisy labels, but also is robust against adversarial perturbations due to its large margin property. Finally, we propose the ensemble version ofDDGC to improve its performance, by investigating the layer-wise characteristics of generative classifier. Our extensive experimental results demonstrate the superiority of DDGC given different learning models optimized by various training techniques to handle noisy labels or adversarial samples. For instance, on CIFAR-10 dataset containing 45% noisy training labels, we improve the test accuracy of a deep model optimized by the state-of-the-art noise-handling training method from33.34% to 43.02%.