Survey on Verifiable Machine Learning

Abstract

Verifiable Machine Learning (VML) provides provable guarantees for ML systems and is becoming a foundation for trust in critical applications. This survey organizes 222 papers published between 2010 and August 2026 along a three-axis taxonomy of verification objectives, technique families, and architectural paradigms. Cryptographic proofs of inference, training, and aggregation account for roughly half of the corpus; robustness and safety verification through formal methods accounts for another large fraction; fairness, unlearning, and ownership form smaller but rapidly growing branches. The survey highlights cross-cutting trust assumptions, scalability versus rigor trade-offs, and deployment barriers, and outlines future directions for trustworthy ML.