Reproduce “Verifying Safety of Neural Networks from Topological Perspectives” 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

Zhen Liang, Dejin Ren, Bai Xue, Ji Wang, Wenjing Yang, Wanwei Liu

Researchers from the National University of Defense Technology and Chinese Academy of Sciences developed a set-boundary reachability method for neural network safety verification, leveraging topological properties like homeomorphisms and open maps. This approach reduced verification computational burden by up to 72% and produced tighter output reachable sets, shrinking interval ranges by 3% to 80% across various network architectures.

Key insights

  • For neural networks exhibiting homeomorphism, analyzing only the input set's boundary is sufficient to verify safety properties for the entire input set, drastically reducing computational effort.
  • The less strict open map property is widely applicable to various NN architectures and also supports effective boundary-based reachability analysis.
  • Leveraging topological properties provides a fundamental approach to reduce the verification domain, directly combating the 'wrapping effect' and high computational costs in NN verification.

Summary by alphaXiv.

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