Eager to Stop: Efficient Falsification of Deep Neural Networks

Guanqin Zhang

Abstract: Deep Neural Networks (DNNs), extensively used in safety-critical domains, require methods to detect misbehavior and ensure provable specifications. DNN testing encounters limitations in time and coverage, affecting effectiveness. DNN verification divides into exact and approximated approaches. Due to scalability challenges, exact methods yield precise outcomes but are suitable for smaller networks. Approximated techniques using abstractions tend to be over-approximated for soundness. Over-approximated verifiers might produce more misleading counterexamples than actual violations, impacting the identification of flaws. This paper proposes a falsifier to efficiently identify counterexamples for DNN robustness by refuting specifications. The proposed approach gradient information to fast approach local optima against specifications, collecting relevant counterexamples effectively.