Verifying Global Two-Safety Properties in Neural Networks with Confidence

Abstract We present the first automated verification technique for confidence-based 2-safety properties, such as global robustness and global fairness, in deep neural networks (DNNs). Our approach combines self-composition to leverage existing reachability analysis techniques and a novel abstraction of the softmax function, which is amenable to automated verification. We characterize and prove the soundness of our static analysis technique. Furthermore, we implement it on top of Marabou, a safety analysis tool for neural networks, conducting a performance evaluation on several publicly available benchmarks for DNN verification.