Logic Gate Neural Networks are Good for Verification

TL;DR

This paper introduces Logic Gate Networks (LGNs), which replace neural network multiplications with Boolean logic gates, making them more suitable for formal verification while maintaining strong predictive capabilities.

Contribution

The paper presents a SAT encoding method for verifying robustness and fairness in LGNs, demonstrating their verification-friendliness and competitive performance on benchmark datasets.

Findings

LGNs are more amenable to symbolic verification.

LGNs maintain strong predictive performance.

Verification method is effective on multiple datasets.

Abstract

Learning-based systems are increasingly deployed across various domains, yet the complexity of traditional neural networks poses significant challenges for formal verification. Unlike conventional neural networks, learned Logic Gate Networks (LGNs) replace multiplications with Boolean logic gates, yielding a sparse, netlist-like architecture that is inherently more amenable to symbolic verification, while still delivering promising performance. In this paper, we introduce a SAT encoding for verifying global robustness and fairness in LGNs. We evaluate our method on five benchmark datasets, including a newly constructed 5-class variant, and find that LGNs are both verification-friendly and maintain strong predictive performance.