Iteratively Enhanced Semidefinite Relaxations for Efficient Neural Network Verification

We propose an enhanced semidefinite program (SDP) relaxation to enable the tight and efficient verification of neural networks (NNs). The tightness improvement is achieved by introducing a nonlinear constraint existing SDP relaxations previously proposed for NN verification. efficiency proposal stems from iterative nature algorithm in that it solves resulting non-convex recursively solving auxiliary convex layer-based problems. show formally solution generated our tighter than state-of-the-art SDP-based solutions problem. also sequence converges optimal relaxation. experimental results on standard benchmarks area achieves performance whilst maintaining acceptable computational cost.