Reproduce “PEREGRiNN: Penalized-Relaxation Greedy Neural Network Verifier” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Haitham Khedr, James Ferlez, Yasser Shoukry

PEREGRiNN, developed by researchers at the University of California, Irvine, presents a sound and complete verifier for ReLU Neural Networks that integrates a penalized convex relaxation with a greedy neuron conditioning strategy. This approach enhanced verification speed and property coverage, proving up to 65% more properties and performing 1.25x to 1.85x faster than state-of-the-art verifiers in adversarial robustness tasks.

Key insights

  • Leveraging a convex objective function with a 'sum-of-slacks' penalty significantly improves the guidance derived from the convex solver, leading to more efficient verification than simple feasibility checking.
  • Strategically weighting the penalty to concentrate 'slack' in earlier unconditioned layers, combined with prioritizing conditioning based on maximum slack, is a powerful heuristic for pruning the search space effectively.
  • Integrating a preliminary input sampling step allows for rapid identification of counterexamples in trivially unsafe cases, complementing the rigorous formal verification process.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.