Towards Verifiable AI with Lightweight Cryptographic Proofs of Inference

This paper tackles the problem of verifying that AI models deployed as cloud services are actually running correctly and using the intended model. Rather than using traditional cryptography (which is too slow for large models), the authors propose a faster verification method that samples and checks random parts of the model's computation trace using statistical properties of neural networks. The approach is millions of times faster than existing methods while still catching common cheating strategies, making it practical for real-world deployment scenarios.