New Method Boosts Neural Network Verification Efficiency

Neural network verification, crucial for ensuring AI system reliability, often relies on Branch and Bound (BaB) techniques. BaB works by breaking down complex verification problems into smaller subproblems, represented as a tree structure. The challenge lies in efficiently identifying the "verdict boundaries" within this tree, which separate verified from unverified subproblems. Current BaB methods are often inefficient, as they sequentially solve each subproblem and require costly bounds propagation. This new research proposes a more effective approach that leverages the monotonic nature of problem paths. It allows for the simultaneous splitting of multiple activation functions, such as ReLUs, and employs an exponential search to skip irrelevant subproblems, significantly speeding up the boundary identification process. An enhanced version further refines this by using quantitative information to estimate boundary positions.