FedBGS: A Blockchain Approach to Segment Gossip Learning in Decentralized Systems

TL;DR

This work addresses privacy-preserving decentralized learning in the presence of non-IID data and the single-point-of-failure risk of centralized federated setups. It introduces FedBGS, a two-phase framework that combines One-Shot Federated K-Means clustering (Federated Analytics) with blockchain-powered segmented gossip learning, storing weights off-chain on IPFS and centroids on-chain. It employs Differential Privacy, Paillier partial homomorphic encryption, and trimmed-mean aggregation to ensure privacy, security, and Byzantine resilience, while Phase 2 enables asynchronous, segment-based updates and leader-driven global aggregation. The approach demonstrates convergence and robustness on standard datasets and discusses practical aspects like Ethereum gas costs and deployment considerations, highlighting potential for scalable, transparent, and secure decentralized learning.

Abstract

Privacy-Preserving Federated Learning (PPFL) is a Decentralized machine learning paradigm that enables multiple participants to collaboratively train a global model without sharing their data with the integration of cryptographic and privacy-based techniques to enhance the security of the global system. This privacy-oriented approach makes PPFL a highly suitable solution for training shared models in sectors where data privacy is a critical concern. In traditional FL, local models are trained on edge devices, and only model updates are shared with a central server, which aggregates them to improve the global model. However, despite the presence of the aforementioned privacy techniques, in the classical Federated structure, the issue of the server as a single-point-of-failure remains, leading to limitations both in terms of security and scalability. This paper introduces FedBGS, a fully Decentralized Blockchain-based framework that leverages Segmented Gossip Learning through Federated Analytics. The proposed system aims to optimize blockchain usage while providing comprehensive protection against all types of attacks, ensuring both privacy, security and non-IID data handling in Federated environments.

Figures (5)

• Figure 1: Example of an Ethereum Block: In Ethereum blocks, the Merkle root wikipediaMerkleTree is a single hash that summarizes and verifies all transactions in the block, ensuring data integrity. The nonce is a number used in the consensus process to achieve a valid block hash. The previous block hash links the current block to the one before it, creating an immutable chain.
• Figure 2: The block size in IPFS can vary depending on the stored data. Thanks to a hierarchical structure based on hash links, each block points to other linked blocks, promoting redundancy and sharing across nodes. This scalable approach ensures fast and reliable retrieval of content from various locations.
• Figure 3: FedBGS overall architecture with cluster-based segmented gossip learning. Thanks to smart contracts and Ethereum's execution logic, every operation delegated to the blockchain is executed across multiple blocks and validated. The segmentation strategy ensures efficient training on non-IID data by limiting updates to assigned neuron blocks in the final layer.
• Figure 4: In the three line charts above, the average accuracies recorded on all grayscale datasets (MNIST and variants) are reported; from left to right, the results correspond to $\beta$=1.0, $\beta$=0.5, and $\beta$=0.1 (the latter under conditions of high heterogeneity among peers). In the lower section, the respective bar plots are shown (each bar plot corresponds to the line chart above it), representing the final and intermediate values for each individual peer.
• Figure 5: Accuracy trend for each peer on the CIFAR-10 dataset. As in the previous line charts, performances with three different $\beta$ values (1.0, 0.5, 0.1) are highlighted to demonstrate the robustness of FedBGS when using more complex RGB datasets.