Adversities in Abstract Interpretation - Accommodating Robustness by Abstract(…)
summary
Presents a formal framework that defines four concrete notions of classifier robustness via abstract interpretation—concrete, abstract robustness, hierarchy-aware robustness, and coherence—and proves implication relations between them. Extends an existing neural-network abstract analyzer, runs experiments on MNIST and CIFAR‑10 that show coherence admits safe within‑class confusions (e.g., ship ↔ airplane) while rejecting cross‑domain errors, reports acceptable runtime numbers in the paper, and proposes approximate coherence for region‑based, scalable checking.
AI-written summary · about
chapters (7)
- 0:00Adversarial attacks motivation
- 2:51Robustness in classifier models
- 8:37Hierarchy-aware abstraction robustness
- 10:03Coherence as robustness
- 12:11Framework validation experiments
- 14:20Language security connections
- 15:43Future work on coherence
related talks
58mBuilding Trust and Safety in Artificial Intelligence with Abstract Interpretation
01h 24mAdversarial Examples
56mAbstract Domains for Machine Learning Verification
42mExploring Pillars of Trustworthy AI: Robustness and Fairness
11mAbstract Interpretation for Automatic Differentiation
16mVerification of Neural Networks' Global Robustness
40mEvasion, Poisoning, Extraction, and Inference: Tools to Defend and Evaluate
24mRobust Inference and Local Algorithms