Blockchain-Enabled Sharded SplitFed Learning for Secure Distributed AI

Briefing

This research addresses the critical issues of centralized vulnerabilities and scalability limitations inherent in existing SplitFed Learning (SFL) frameworks. It proposes Blockchain-enabled SplitFed Learning (BSFL), a novel architecture that integrates sharding with a committee-driven blockchain consensus mechanism to decentralize operations and enhance security. This foundational breakthrough enables robust, scalable, and secure distributed machine learning, eliminating the need for central trust and mitigating risks such as data poisoning and single points of failure.

Context

Prior to this research, collaborative learning techniques like Federated Learning (FL) and Split Learning (SL) offered solutions for privacy-preserving model training but faced significant limitations. FL imposed substantial computational demands on client devices, while SL suffered from prolonged training times due to sequential interactions. SplitFed Learning (SFL) emerged as a hybrid, yet it inherited critical scalability and security issues from its predecessors, notably the risks associated with centralized servers, which include single points of failure, potential for malicious manipulation, and privacy breaches.

Analysis

The paper’s core mechanism, Blockchain-enabled SplitFed Learning (BSFL), fundamentally transforms SFL by decentralizing its architecture. This is achieved through two primary innovations ∞ sharding and a committee consensus mechanism. Sharding distributes the computational workload of the Split Learning (SL) server across multiple parallel shards, significantly enhancing scalability and efficiency. The blockchain integration replaces the centralized Federated Learning (FL) server, with smart contracts autonomously managing tasks such as model aggregation, update validation, and reward distribution.

A committee-driven consensus mechanism is introduced, where SFL servers within each shard form a committee to validate and score model updates from other members. Only the top-performing models are aggregated, and committee members are dynamically rotated across cycles to prevent collusion and ensure fairness. This approach effectively filters out malicious or tampered model updates, fortifying the system against data poisoning attacks.

Parameters

• Core Concept ∞ Blockchain-enabled SplitFed Learning (BSFL)
• Key Mechanism ∞ Committee Consensus
• Scalability Primitive ∞ Sharding
• Decentralization Tool ∞ Smart Contracts
• Problem Addressed ∞ Centralized Vulnerabilities
• Attack Mitigation ∞ Data Poisoning Filtering
• Performance Metric ∞ Round Completion Time
• Blockchain Platform ∞ Hyperledger Fabric
• Authors ∞ Sokhankhosh, A. et al.
• Dataset Used ∞ Fashion MNIST

This research opens new avenues for highly secure and scalable distributed machine learning. Future work includes extending BSFL to support multi-part model splits for enhanced client-side privacy, which could further reduce data exposure. Additionally, adapting evaluation metrics beyond traditional losses for generative applications will be crucial, offering deeper insights into model performance in diverse scenarios. These advancements promise to unlock more robust and privacy-preserving AI models across various real-world applications within the next three to five years.

Signal Acquired from ∞ arXiv.org