VeRAPAk: Distributed Counterexample Generation via Iterative Neural Network Verification and Falsification (Artifact)

Le résumé fourni par la source

This artifact accompanies the FMCAD 2026 submission, "VeRAPAk: Distributed Counterexample Generation via Iterative Neural Network Verification and Falsification" It provides the complete, Dockerized reproduction package for VeRAPAk, a modular framework for neural network verification and distributed counterexample generation. The repository contains the framework’s source code—including the iterative partitioning engine and the novel Randomized Fast Gradient Sign Method (RFGSM)—alongside the benchmarking scripts necessary to validate the paper's core experimental claims. Note on Hardware Requirements: Execution of this artifact requires an x86_64 CPU with AVX instruction support and at least 32 GB of RAM. M-series (Apple Silicon) Macs are not supported natively due to hardware-specific dependencies in the underlying machine learning libraries. Storage Requirements: Please ensure your system has sufficient disk space prior to initialization. The base Docker image requires approximately 60 GB of storage once loaded. Running the full reproduction of results will generate additional image data, roughly 100 GB. Intermediate results can be removed with provided scripts (see README.md) to aleviate some of this overhead. Please see the included README.md for complete setup instructions, Docker initialization commands, and a detailed breakdown of the expected outputs.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.