UnifyFL proposes a decentralized cross-silo federated learning framework using Ethereum blockchain and IPFS to enable trust-based collaboration among organizations, achieving comparable accuracy to centralized FL with flexible aggregation policies and efficient handling of stragglers through synchronous and asynchronous modes.
Federated Learning, Decentralized Systems, Cross-Silo Collaboration, Trustworthy AI, Efficiency, Scalability
Sarang S, Druva Dhakshinamoorthy, Aditya Shiva Sharma, Yuvraj Singh Bhadauria, Siddharth Chaitra Vivek, Arihant Bansal, Arnab K. Paul
BITS Pilani, KK Birla Goa Campus, India
Generated by grok-3
Background Problem
The paper addresses the challenge of enabling collaboration among organizations in federated learning (FL) while balancing trust and resource efficiency. In traditional FL, organizations often operate independently, limiting model performance due to data silos and heterogeneity. Existing solutions like centralized aggregation introduce trust issues with a single point of failure, while peer-to-peer approaches incur high computational and network overheads. UnifyFL aims to solve these problems by proposing a decentralized cross-silo FL framework that fosters trust and collaboration without relying on a central aggregator, thus enhancing privacy, security, and resource utilization.
Method
UnifyFL introduces a trust-based cross-silo FL framework using decentralized orchestration and distributed storage. Its core idea is to eliminate the need for a central aggregator by leveraging a private Ethereum blockchain as a decentralized orchestrator and InterPlanetary File System (IPFS) for distributed storage of model weights. The workflow involves: (1) local model aggregation by clusters, storing weights on IPFS, and submitting content identifiers (CIDs) to the blockchain; (2) random selection of a majority of scorers to evaluate models using accuracy or similarity-based scoring; (3) aggregators querying the blockchain for scored models and applying customizable aggregation policies (e.g., Top k, Random k) to update global models. It operates in synchronous (Sync) mode, where all clusters train and score together, and asynchronous (Async) mode, allowing independent operation to handle stragglers. Smart contracts on Ethereum manage coordination, ensuring transparency and immutability.
Experiment
The experiments were conducted on two testbeds: a high-performance GPU cluster (4 nodes with NVIDIA RTX A2000) for Tiny ImageNet with VGG16 (138M parameters), and a resource-constrained edge cluster (Raspberry Pi, Jetson Nano) for CIFAR-10 with a lightweight CNN (62K parameters). Datasets were partitioned as IID and non-IID (Dirichlet with α=0.1, 0.5) to simulate realistic heterogeneity. The setup compared UnifyFL against a centralized hierarchical FL baseline (HBFL), testing Sync and Async modes, various aggregation policies (e.g., Pick All, Top k), and scalability up to 60 clients. Results showed UnifyFL achieves comparable accuracy (e.g., 35-38% on Tiny ImageNet vs. 36.8% baseline) with faster convergence in Async mode (4000s vs. 6000s in Sync). The design handles device heterogeneity well, with minimal overhead (Geth: 0.2% CPU, IPFS: 3.5% CPU). However, Async mode sometimes resulted in lower accuracy due to limited model availability, and testing on standard datasets may not fully reflect real-world cross-silo challenges. The experimental setup is comprehensive for initial validation but lacks stress testing under extreme conditions or advanced privacy attacks.
Further Thoughts
UnifyFL’s approach to decentralized FL using blockchain and IPFS is a step forward in addressing trust issues, but it raises questions about practical deployment in latency-sensitive applications due to blockchain transaction delays, even in a private setting. A deeper exploration could involve hybrid approaches combining blockchain with faster consensus mechanisms like Raft or Paxos for orchestration. Additionally, the paper’s focus on cross-silo FL could be extended to cross-device scenarios with extreme heterogeneity, potentially integrating with emerging edge computing paradigms. The lack of privacy-preserving techniques like differential privacy is a notable gap; future work could draw inspiration from frameworks like CrypTFlow or SecureML to enhance data security. Moreover, testing UnifyFL in domains like healthcare or finance, where data sensitivity and regulatory constraints are paramount, could reveal unique challenges and validate its real-world applicability. Finally, the flexibility in aggregation policies, while innovative, might benefit from a meta-learning layer to dynamically optimize policy selection across clusters, drawing from recent advances in adaptive FL systems.