July 16, 2026 Tight Certification of Adversarially Trained Neural Networks via Nonconvex Low-Rank Semidefinite Relaxations 基于低秩约束 SDP 松弛的非凸方法，突破 LP 验证器的凸松弛屏障 neural-network-verification adversarial-robustness