Differential Verification of Deep Neural Networks
summary
Presents an abstract-interpretation technique for verifying differential properties of deep neural networks—properties that compare two networks such as network equivalence, input sensitivity, and stability. Uses sound abstractions that relate intermediate values of two structurally similar networks to tightly bound their maximum difference over all inputs, and includes automated synthesis methods to create linear abstractions for arbitrary nonlinear functions so the approach applies beyond feed-forward ReLU architectures.
AI-written summary · about
chapters (19)
- 0:00Start
- 1:50Example: Neural Network Compression
- 10:11Limitation: Generality
- 14:47Neural Network Verification
- 16:11A Toy Example
- 17:16Baseline Differential Verification
- 21:38Baseline Approach
- 29:37Summary
- 35:32Evaluation
- 39:10Previous Approach
- 40:48NeuroDiff Approach
- 42:24Recap
- 44:48Defining a Good Approximation
- 45:24Existing Approach: Decomposing
- 46:47Proposed Online Synthesis
- 47:24Relax to Linear Optimization
- 47:58Verifying the Approximation
- 51:56Outline
- 53:13Motivating Example
related talks
17mInput-Relational Verification of Deep Neural Networks
42mTowards Verification of General Neural Networks: A Dual Approach
24mVerification of Neural Networks’ Local Differential Classification Privacy
55mTowards Verified Deep Learning
20mIncremental Verification of Neural Networks
16mVerification of Neural Networks' Global Robustness
59mWhats and Whys of Neural Network Verification (A Declarative Programming Perspec...
56mAbstract Domains for Machine Learning Verification