This repository implements an end-to-end pipeline for robustness verification of binary neural networks (BNNs) by mapping the verification problem to a Quadratic Unconstrained Binary Optimization (QUBO) instance. Starting from a trained ten-class BNN and a correctly classified input, converts it into a dense QUBO using penalty methods, and searches for adversarial perturbations within a prescribed budget.
The framework is designed to work with both conventional algorithms and unconventional Ising-style hardware.
The goal of this codebase is demonstrate that BNN robustness verfication problem can be expressed as QUBO instances which cab solved on both classical and emerging Ising/annealing platforms, providing a bridge between AI trustworthiness and unconventional computing.