Approximate Conformance Verification of Deep Neural Networks

We consider the problem of approximate conformance checking on deep neural networks. More precisely, given two neural networks and a conformance bound epsilon, we need to check if the neural network outputs are within epsilon given the same inputs from the input set. Our approach reduces the approximate conformance checking problem to a reachability analysis problem using transformations of neural networks. We provide experimental comparison of epsilon-conformance checking based on our approach using various reachability analysis tools as well as other alternate epsilon-conformance checking algorithms. We illustrate the benefits of our approach as well as identify reachability analysis tools that are conducive for conformance checking.