Verification of Neural Networks' Global Robustness
summary
Introduces Delta Global robustness as the minimal confidence margin guaranteeing any input stays correctly classified under a given perturbation, and builds an anytime verifier VHAGaR that encodes two network copies into mixed-integer programming, prunes booleans by encoding cross-network dependencies, and injects a hyper-adversarial attack to seed tight lower bounds. On FashionMNIST/CIFAR-10 with a 3-hour timeout VHAGaR gives average lower–upper gap 1.9 vs 154.7 for a prior verifier, runs 130.6× faster, and shows a 78.6× speedup from dependencies plus attacks.
AI-written summary · about
chapters (15)
- 0:00Neural network robustness intro
- 1:26Local robustness limits
- 2:10Global robustness guarantees
- 3:55Delta Global robustness
- 5:21Formal Delta Global definition
- 6:43Problem to optimization
- 8:08Mixed integer programming
- 8:49Anytime optimization
- 9:29Network dependency encoding
- 11:15Hyper adversarial attack
- 12:39Vager system overview
- 13:24Vager formal guarantees
- 13:45Baseline method evaluation
- 14:25Robustness heatmap insights
- 15:07Conclusion and future
related talks
42mTowards Verification of General Neural Networks: A Dual Approach
20mIncremental Verification of Neural Networks
11mExploiting Verified Neural Networks via Floating Point Numerical Error
17mMini-Batch Robustness Verification of Deep Neural Networks
17mInput-Relational Verification of Deep Neural Networks
42mExploring Pillars of Trustworthy AI: Robustness and Fairness
01h 24mAdversarial Examples
55mTowards Verified Deep Learning