Error-Correcting Output Codes with Ensemble Diversity for Robust Learning in Neural Networks

Though deep learning has been applied successfully in many scenarios, malicious inputs with human-imperceptible perturbations can make it vulnerable real applications. This paper proposes an error-correcting neural network (ECNN) that combines a set of binary classifiers to combat adversarial examples the multi-class classification problem. To build ECNN, we propose design code matrix so minimum Hamming distance between any two rows (i.e., codewords) and shared information columns partitions class labels) are simultaneously maximized. Maximizing row distances increase system fault tolerance while maximizing column helps diversity classifiers. We end-to-end training method for our which allows further improvement The renders proposed ECNN different from traditional output (ECOC) based methods train independently. is complementary other existing defense approaches such as be conjunction them. empirically demonstrate effective against state-of-the-art white-box black-box attacks on several datasets maintaining good accuracy normal examples.