Blockchain-Based Decentralized Model Aggregation for Cross-Silo Federated Learning in Industry 4.0

Abstract

Traditional federated learning (FL) adopts a client-server architecture where FL clients (e.g., IoT edge devices) train a common global model with the help of a centralized orchestrator (cloud server). However, current approaches are moving away from centralized orchestration toward a decentralized one in order to fully adapt FL for a cross-silo configuration with multiple organizations acting as clients. State-of-the-art decentralized FL mechanisms make at least one of the following assumptions: 1) clients are trusted organizations and cannot inject low-quality model updates for aggregation and 2) client local models can be shared with other clients or a third party for verification of low-quality updates. This article proposes a Blockchain-based decentralized framework for scenarios where participatory organizations are believed to be fully capable of injecting low-quality model updates as they are not willing to expose their local models to any other entity for verification purpose. The proposed decentralized FL framework adopts a novel hierarchical network of aggregators with the ability to punish/reward organizations in proportion to their local model quality updates. The framework is flexible and unlike state-of-the-art solutions, prevents a single entity from possessing the aggregated model in any FL round of training. The proposed framework is tested with respect to off-chain and on-chain performance in two Industry 4.0 use cases: 1) predictive maintenance and 2) product visual inspection. A comparative evaluation against the state-of-the-art reveals the proposed framework's utility in terms of minimizing model convergence time and latency while maximizing accuracy and throughput.