Year: 2026
Type: dissertation
Abstract: Federated learning (FL) enables collaborative model training across distributed IoT devices without sharing raw data, but introduces a fundamental verification gap: malicious clients can submit fabricated gradient updates to corrupt the global model without the server being able to verify honest computation. Existing zero-knowledge proof (ZKP) defenses provide strong guarantees but impose proof-to-training time (PT/TT) ratios of 30--40, making deployment on resource-constrained IoT hardware infeasible. This thesis investigates whether partial ZKP verification, proving only a subset of the training computation, provides a verifiable algebraic integrity check at IoT-compatible cost. We implement a ZKP-FL system in Noir and Barretenberg UltraHonk, consisting of three circuits: a last-layer outer product circuit (FC3, 840 constraints), an optional second-to-last-layer circuit (FC2, 10,920 constraints), and a Pedersen hash preimage circuit for client identity verification. The system is evaluated on a LeNet-5 trained on MNIST with 5 federated clients across 5 rounds at 12,000 samples per client. Results show that the outer product circuit rejects all three evaluated submission strategies across all rounds and sample sizes, with a PT/TT ratio of 0.001, three to four orders of magnitude below the state of the art, while proof generation time remains constant regardless of training data volume. These findings demonstrate that partial layer verification is integratable at IoT-compatible overhead and that ZKP overhead can be decoupled from training data volume in federated learning. The formal characterisation of what adversary classes the algebraic check provably resists is identified as the primary open question for future work.
Author: Martinez Sanchez, Hector Miguel
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7221496201