Abstract When validated neural networks (NNs) are pruned (and retrained) before deployment, it is desirable to prove that the new NN behaves equivalently to the (original) reference NN. To this end, our paper revisits the idea of differential verification which performs reasoning on differences between NNs: On the one hand, our paper proposes a novel abstract domain for differential verification admitting more efficient reasoning about equivalence. On the other hand, we investigate empirically and theoretically which equivalence properties are (not) efficiently solved using differential reasoning. Based on the gained insights, and following a recent line of work on confidence-based verification, we propose a novel equivalence property that is amenable to Differential Verification while providing guarantees for large parts of the input space instead of small-scale guarantees constructed w.r.t. predetermined input points. To this end, we propose an improved approach for (approximately) constraining an NN’s softmax confidence. We implement our approach in a new tool called VeryDiff and perform an extensive evaluation on numerous old and new benchmark families, including new pruned NNs for particle jet classification in the context of CERN’s LHC where we observe median speedups $$>300\times $$ > 300 × over the State-of-the-Art verifier $$\alpha ,\beta $$ α , β -CROWN.

As neural networks (NNs) are increasingly introduced into safety-critical domains, there is a growing need to formally verify NNs before deployment. In this work we focus on the formal verification problem of NN equivalence which aims to prove that two NNs (e.g. an original and a compressed version) show equivalent behavior. Two approaches have been proposed for this problem: Mixed integer linear programming and interval propagation. While the first approach lacks scalability, the latter is only suitable for structurally similar NNs with small weight changes.The contribution of our paper has four parts. First, we show a theoretical result by proving that the epsilon-equivalence problem is coNP-complete. Secondly, we extend Tran et al.’s single NN geometric path enumeration algorithm to a setting with multiple NNs. In a third step, we implement the extended algorithm for equivalence verification and evaluate optimizations necessary for its practical use. Finally, we perform a comparative evaluation showing use-cases where our approach outperforms the previous state of the art, both, for equivalence verification as well as for counter-example finding.