<ul><li>The increasing use of artificial intelligence in critical applications requires effective neural network certification, particularly against 'patch attacks' that obscure parts of images.</li><li>Current preimage approximation methods, like the PREMAP algorithm, face scalability challenges in certification against adversarial inputs.</li><li>This paper introduces algorithmic enhancements to the PREMAP algorithm, including tighter bounds, adaptive Monte Carlo sampling, and improved branching heuristics.</li><li>The improved method shows significant efficiency gains in reinforcement learning control benchmarks and scales to previously challenging convolutional neural networks, highlighting its potential for reliability and robustness certification.</li></ul>

Efficient Preimage Approximation for Neural Network Certification

Discover more