A Guide to NNV¶
Welcome to the Guide of Neural Network Verification (NNV). This comprehensive collection of technical guides covers foundational concepts to cutting-edge research techniques, designed for practitioners and researchers interested in formal verification, robustness testing, and trustworthy AI.
The guides are organized into four complementary phases. Click on any phase below to explore the topics.
Understanding the verification challenge: motivation, threat models, mathematical formulation, soundness, completeness, systematic taxonomy, and theoretical barriers.
Methods for certifying robustness: decomposition framework, activation handling, bound propagation, LP/SDP relaxations, Lipschitz methods, and complete verification (SMT, MILP, branch-and-bound).
Building and evaluating robust networks: formal specifications, verification workflows, benchmarks, robustness testing, regularization training, and certified adversarial training.
Research directions and real-world deployment: advanced training strategies, certified defenses, scalability challenges, alternative threats, diverse architectures, and real-world applications.