Probabilistic Verification of Neural Networks via Hull Generation

A novel framework has been developed for the probabilistic verification of neural networks, which calculates a guaranteed range for safe probabilities by effectively identifying both safe and unsafe probabilistic hulls. This method employs a strategy of subdividing the state space using regression trees, incorporates a boundary-aware sampling technique to pinpoint safety boundaries, and utilizes iterative refinement through probabilistic prioritization. This innovative approach tackles the challenge of ensuring neural network safety in the presence of probabilistic disturbances in inputs.