The Convex Relaxation Barrier, Revisited: Tightened Single-Neuron Relaxations for Neural Network Verification

A fundamental problem in deep neural networks is to verify or certify that a trained network is robust, i.e. not susceptible to adversarial attacks [