Zero-Knowledge Proofs for Machine Learning and Compound AI Systems: A Survey

PhD Qualifying Examination


Title: "Zero-Knowledge Proofs for Machine Learning and Compound AI
Systems: A Survey"

by

Mr. Zhaoyu WANG


Abstract:

Machine learning (ML) models are increasingly deployed as services whose
users observe only an output. The training data is confidential, the model
parameters are proprietary, and the computation runs on infrastructure the
user does not control, so nothing in the returned result reveals whether it
was produced as claimed. Zero-knowledge proofs (ZKPs), cryptographic
protocols that allow a prover to convince a verifier that a computation
was performed correctly while revealing nothing beyond its output, are
considered a promising solution to this integrity problem. Their practical
adoption in ML, however, faces significant obstacles. Model arithmetic is
real-valued and nonlinear, whereas proof systems support only finite
fields and low-degree constraints. The control flow around a deployed model
depends on its data, while a circuit must be fixed before that data is
known. Compounding both, proving remains orders of magnitude more
expensive than computing. To overcome these hurdles, recent research
focuses on reformulating ML computations into structures that proof
systems can evaluate efficiently.

In this survey, we present a comprehensive review of ZKP-based verifiable
ML, organized by what is being verified. We begin with the fundamental
concepts of ZKPs and the hurdles that arise when applying them to ML
workloads. We then examine proofs of inference, and proofs of training in
both the single-prover and the federated setting. Proofs of model
properties follow, covering claims such as fairness and privacy compliance
that no single execution witnesses. Finally, we discuss compound AI
systems, where the verification target spans retrieval, agent execution,
and multi-stage analytical pipelines. By synthesizing current knowledge,
this survey aims to provide practical guidance for applying ZKPs in ML
systems and to identify the directions in which further progress is most
needed.


Date:                   Tuesday, 18 August 2026

Time:                   2:00pm - 4:00pm

Venue:                  Room 5506
                        Lift 25/26

Committee Members:      Dr. Shuai Wang (Supervisor)
                        Prof. Ke Yi (Chairperson)
                        Dr. Mingxun Zhou