Paper 2025/1732 Original: eprint.iacr.org/2025/1732 ↗

Zero-Knowledge AI Inference with High Precision

Abstract

Artificial Intelligence as a Service (AIaaS) enables users to query a model hosted by a service provider and receive inference results from a pre-trained model. Although AIaaS makes artificial intelligence more accessible, particularly for resource-limited users, it also raises verifiability and privacy concerns for the client and server, respectively. While zero-knowledge proof techniques can address these concerns simultaneously, they incur high proving costs due to the non-linear operations involved in AI inference and suffer from precision loss because they rely on fixed-point representations to model real numbers. In this work, we present ZIP, an efficient and precise commit and prove zero-knowledge SNARK for AIaaS inference (both linear and non-linear layers) that natively supports IEEE-754 double-precision floating-point semantics while addressing reliability and privacy challenges inherent in AIaaS. At its core, ZIP introduces a novel relative-error-driven technique that efficiently proves the correctness of complex non-linear layers in AI inference computations without any loss of precision, and hardens existing lookup-table and range proofs with novel arithmetic constraints to defend against malicious provers. We implement ZIP and evaluate it on standard datasets (e.g., MNIST, UTKFace, and SST-2). Our experimental results show, for non-linear activation functions, ZIP reduces circuit size by up to three orders of magnitude while maintaining the full precision required by modern AI workloads.