The CAISAR Platform: Extending the Reach of Machine Learning Specification and Verification

Michele Alberti, François Bobot, Julien Girard-Satabin, Alban Grastien, Aymeric Varasse, Zakaria Chihani

Abstract: The formal specification and verification of machine learning models have advanced remarkably in less than a decade, leading to a profusion of verification tools that provide mathematical guarantees about model properties. However, this growing diversity risks ecosystem fragmentation, making it difficult to compare tools beyond narrowly defined benchmarks. Moreover, much of the progress to date has focused on a limited class of properties, particularly local robustness. While existing tools are increasingly effective at verifying such properties, more complex ones, such as those involving multiple neural networks, remain beyond their capabilities: these properties cannot currently be expressed in their specification languages, nor can they be directly verified. This applies even to the winning verification tools of the International Verification of Neural Networks Competition (VNN-Comp).

External IDs:doi:10.1007/978-3-032-10794-7_15