Abstract
The safety-critical scenarios of artificial intelligence (AI), such as autonomous driving, Internet of Things, and smart healthcare, have raised critical requirements for trustworthy AI to guarantee privacy and security with reliable decisions. As a nascent branch for trustworthy AI, federated learning (FL) has been regarded as a promising privacy-preserving framework for training a global AI model over collaborative devices. However, security challenges still exist, such as Byzantine attacks from malicious devices and model tampering from malicious servers. In this paper, we propose a decentralized blockchain-based FL (B-FL) architecture using a secure global aggregation algorithm and a practical Byzantine fault tolerance (PBFT) consensus protocol among multiple edge servers. To address the training latency induced by blockchain consensus, we formulate a network optimization problem for bandwidth and power allocation and solve it using a deep reinforcement learning (DRL) based algorithm. Simulation results demonstrate that B-FL resists malicious attacks while significantly reducing training latency compared to baseline algorithms.
Introduction
Federated Learning (FL) is the primary framework for privacy-preserving AI, allowing devices to train a global model without sharing raw data. However, it suffers from two major trust issues: malicious devices uploading poisoned models (Byzantine attacks) and centralized servers that represent a single point of failure or a source of model tampering. While blockchain can decentralize this process, it often introduces significant latency that is impractical for edge deployment. This paper introduces a system that is both secure and optimized for real-world edge performance.
The B-FL Architecture
The proposed B-FL architecture replaces the single central server with a network of edge servers, including a primary server and several validators. This structure creates a decentralized environment where model updates are verified across multiple nodes before being committed to the blockchain.
PBFT Consensus and Secure Aggregation
To ensure integrity without the high energy cost of Proof-of-Work, the system utilizes the Practical Byzantine Fault Tolerance (PBFT) consensus protocol. This permissioned protocol requires only a two-thirds majority to validate updates, making it suitable for edge servers.
The process follows a structured workflow:
- Local Training: Edge devices train models on local datasets.
- Upload & Aggregation: Devices upload local models to the Primary Edge Server, which uses the
multi-KRUMalgorithm to filter out malicious outliers based on Euclidean distances. - Consensus Phases: The servers execute Pre-Prepare, Prepare, and Commit steps to reach a consensus on the aggregated model.
- Global Update: Once validated, the global model is stored in a block and downloaded by devices for the next training round.
Resource Allocation via Deep Reinforcement Learning
To tackle the latency challenges of blockchain consensus, the authors optimize bandwidth and power allocation. This is modeled as a Markov Decision Process (MDP) and solved using the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm. TD3 is particularly effective at handling the continuous variables inherent in network resource management, minimizing long-term average training latency across progressive learning rounds.
Key Results and Evaluation
The system was tested on MNIST, CIFAR-10, and a Heart Activity dataset (representing non-IID IoT data).
- Attack Resistance: B-FL maintained a steady 97.9% accuracy on MNIST even when 40% of devices were malicious, whereas standard FL performance collapsed.
- Latency Optimization: The TD3-based resource allocation significantly outperformed random and average allocation baselines, maintaining low latency even as the number of devices scaled.
- Convergence: The DRL agent proved stable, with rewards converging quickly, demonstrating its ability to adapt to dynamic wireless channel conditions.
Conclusion
The B-FL framework successfully combines blockchain-based security with DRL-driven efficiency. By addressing both device-side poisoning and server-side tampering while optimizing for edge latency, it provides a viable path for trustworthy AI in safety-critical applications. Future research will focus on adapting this architecture for larger, more dynamic network scales and permissionless environments.
Summary updated