Neural Network Verification Using Residual Reasoning

With the increasing integration of neural networks as components in mission-critical systems, there is an need to ensure that they satisfy various safety and liveness requirements. In recent years, numerous sound complete verification methods have been proposed towards end, but these typically suffer from severe scalability limitations. Recent work has enhancing such techniques with abstraction-refinement capabilities, which shown boost scalability: instead verifying a large complex network, verifier constructs then verifies much smaller whose correctness implies original network. A shortcoming scheme if network fails, needs perform refinement step increases size being verified, start new scratch—effectively “wasting” its earlier on this paper, we present enhancement abstraction-based networks, by using residual reasoning: process utilizing information acquired when abstract order expedite refined essence, method allows store about parts search space guaranteed behave correctly, it focus areas where bugs might be discovered. We implemented our approach extension Marabou verifier, obtained promising results.