Partial Model Sharing Improves Byzantine Resilience in Federated Conformal Prediction

2026-05-12

AI Summary

This work addresses the challenge of Byzantine attacks that simultaneously compromise both training and calibration phases in federated conformal prediction by proposing the first unified defense framework. To reduce attack surface and communication overhead, clients exchange only partial model parameters and compress their nonconformity scores into histogram-based feature vectors. The server then employs distance-based metrics on these features to detect and exclude malicious participants, estimating conformal quantiles exclusively from benign data. The proposed method achieves coverage close to the nominal level under various Byzantine attacks while yielding substantially tighter prediction intervals, thereby enhancing both robustness and communication efficiency.

Abstract

We propose a Byzantine-resilient federated conformal prediction (FCP) method that leverages partial model sharing, where only a subset of model parameters is exchanged each round. Unlike existing robust FCP approaches that primarily harden the calibration stage, our method protects both the federated training and conformal calibration phases. During training, partial sharing inherently restricts the attack surface and attenuates poisoned updates while reducing communication. During calibration, clients compress their non-conformity scores into histogram-based characterization vectors, enabling the server to detect Byzantine clients via distance-based maliciousness scores and to estimate the conformal quantile using only benign contributors. Experiments across diverse Byzantine attack scenarios show that the proposed method achieves closer-to-nominal coverage with substantially tighter prediction intervals than standard FCP, establishing a robust and communication-efficient approach to federated uncertainty quantification.

Problem

• Byzantine resilience
• federated conformal prediction
• partial model sharing
• uncertainty quantification
• malicious client detection

Innovation

• partial model sharing
• Byzantine resilience
• federated conformal prediction
• histogram-based characterization
• maliciousness detection