Correct orchestration of Federated Learning generic algorithms: formalisation and verification in CSP

TL;DR

This paper formalizes and verifies the correctness of centralized and decentralized federated learning algorithms using CSP process calculus and the PAT model checker, ensuring deadlock freedom and successful termination.

Contribution

It introduces a formal verification approach for generic federated learning algorithms, bridging real Python implementations with CSP models and automatic correctness proofs.

Findings

Verified deadlock freeness of FL algorithms

Proved successful termination of FL algorithms

Established formal correctness framework for FL orchestration

Abstract

Federated learning (FL) is a machine learning setting where clients keep the training data decentralised and collaboratively train a model either under the coordination of a central server (centralised FL) or in a peer-to-peer network (decentralised FL). Correct orchestration is one of the main challenges. In this paper, we formally verify the correctness of two generic FL algorithms, a centralised and a decentralised one, using the CSP process calculus and the PAT model checker. The CSP models consist of CSP processes corresponding to generic FL algorithm instances. PAT automatically proves the correctness of the two generic FL algorithms by proving their deadlock freeness (safety property) and successful termination (liveness property). The CSP models are constructed bottom-up by hand as a faithful representation of the real Python code and is automatically checked top-down by PAT.