Robust-by-Design Classification via Unitary-Gradient Neural Networks

The use of neural networks in safety-critical systems requires safe and robust models, due to the existence adversarial attacks. Knowing minimal perturbation any input x, or, equivalently, knowing distance x from classification boundary, allows evaluating robustness, providing certifiable predictions. Unfortunately, state-of-the-art techniques for computing such a are computationally expensive hence not suited online applications. This work proposes novel family classifiers, namely Signed Distance Classifiers (SDCs), that, theoretical perspective, directly output exact rather than probability score (e.g., SoftMax). SDCs represent robust-by-design classifiers. To practically address requirements an SDC, network architecture named Unitary-Gradient Neural Network is presented. Experimental results show that proposed approximates signed classifier, allowing at cost single inference.