REGLO: Provable Neural Network Repair for Global Robustness Properties

• DOI: 10.1609/aaai.v38i11.29094
• Publication Date: 2024-03-25T10:59:35Z
• Authors: Feisi Fu; Zhilu Wang; Weichao Zhou; Yixuan Wang; Jiameng Fan; Chao Huang; Qi Zhu; Xin Chen; Wenchao Li
• Keywords: Robustness

We present REGLO, a novel methodology for repairing pretrained neural networks to satisfy global robustness and individual fairness properties. A network is said be globally robust with respect given input region if only all the points in are locally robust. This notion of also captures as special case. prove that any counterexample property must exhibit corresponding large gradient. For ReLU networks, this result allows us efficiently identify linear regions violate property. By formulating solving suitable convex optimization problem, REGLO then computes minimal weight change will provably repair these violating regions.