Phase 2: Core Verification Techniques¶
Methods for Certifying Robustness
1 min read · 288 words
Building on the foundations, this phase provides comprehensive coverage of the core algorithmic approaches to neural network verification. You’ll learn the unary-binary decomposition framework that underlies all methods, understand how activation functions affect verification complexity, master bound propagation techniques, explore incomplete methods (LP relaxations, multi-neuron relaxations, SDP, Lipschitz bounds), and study complete verification approaches (SMT, MILP, branch-and-bound). This phase focuses on the technical depth of verification algorithms.
What You’ll Learn
This phase covers core verification algorithms: from the fundamental decomposition framework and bound propagation to incomplete methods (LP, SDP, Lipschitz) and complete approaches (SMT, MILP, branch-and-bound).
Guides in This Phase¶
The unary-binary framework foundation for all verification methods
How different activations affect verification and their approximation techniques
IBP, CROWN, DeepPoly, and other core incomplete verification techniques
LP relaxation, dual formulation, and triangle relaxation
Capturing neuron correlations with product-based relaxations
Semi-definite programming relaxations and nuclear norm approaches
Lipschitz constant estimation and spectral normalization
Constraint solving with SAT/SMT and mixed-integer linear programming
Specialized complete verification using the Reluplex algorithm
Hybrid approaches combining branching with bound propagation (α-β-CROWN)
Next Phase
After mastering the core techniques, move on to Phase 3: Robust Training & Practical Implementation to learn about robust training and practical implementation.
Comments & Discussion