Source-linked AI summary
A Blockchain-based Decentralized Federated Learning Framework with Committee Consensus
Yuzheng Li, Chuan Chen, Nan Liu, Huawei Huang, Zibin Zheng, Qiang Yan
TL;DR
Federated learning must protect private data while addressing attacks by malicious clients or central servers. BFLC uses blockchain-based decentralized storage and committee consensus to manage models and updates, and the paper reports effectiveness on real-world data and resistance to malicious attacks.
Problem
Federated learning faces malicious clients and central servers that can damage global models or expose users’ private data.
Method
BFLC uses blockchain to store global models and local updates, with committee consensus validating updates and generating blocks.
Results
Experiments on real-world federated-learning data verify BFLC effectiveness, while simulations show resistance to higher malicious-node proportions than compared methods.
Takeaways & Limitations
BFLC provides a decentralized federated-learning framework whose committee mechanism reduces consensus computation and mitigates malicious updates.
Abstract
from arXiv · showhide
Federated learning has been widely studied and applied to various scenarios. In mobile computing scenarios, federated learning protects users from exposing their private data, while cooperatively training the global model for a variety of real-world applications. However, the security of federated learning is increasingly being questioned, due to the malicious clients or central servers' constant attack to the global model or user privacy data. To address these security issues, we proposed a decentralized federated learning framework based on blockchain, i.e., a Blockchain-based Federated Learning framework with Committee consensus (BFLC). The framework uses blockchain for the global model storage and the local model update exchange. To enable the proposed BFLC, we also devised an innovative committee consensus mechanism, which can effectively reduce the amount of consensus computing and reduce malicious attacks. We then discussed the scalability of BFLC, including theoretical security, storage optimization, and incentives. Finally, we performed experiments using real-world datasets to verify the effectiveness of the BFLC framework.
I. INTRODUCTION
The paper identifies security, centralization, consensus efficiency, and storage challenges in blockchain-based federated learning, then proposes BFLC to address them. BFLC combines decentralized model storage, committee validation, storage optimization, and incentive-related scalability considerations.
- Motivation: Centralized FL servers create bandwidth, stability, fairness, model-poisoning, and privacy risks.The server aggregates updates, selects clients, maintains the global model, and broadcasts models to participating nodes.
- Motivation: Blockchain can replace the central server, but decentralized FL still faces consensus-efficiency and blockchain-storage burdens.The paper also notes that transferring aggregation to clients can increase computation and network pressure on nodes.
- Proposed solution: The committee validates updates, records qualified updates on the blockchain, and enables honest nodes to reinforce one another while ignoring some malicious updates.Historical blocks can be abandoned by partial nodes to reduce storage use, while contributing nodes may receive rewards through incentives.
- Proposed solution: BFLC defines decentralized model and update storage, a training process, and a committee consensus mechanism.The proposed architecture uses an alliance chain for node permission control and a storage pattern intended to provide rapid access to the latest model.
- Evaluation: Experiments on a real-world federated-learning dataset evaluate BFLC effectiveness, while simulations examine security against malicious attacks.The paper separately discusses scalability through node management, malicious-node analysis, and storage optimization.
II. RELATED WORK
Prior work applies decentralization and blockchain to federated learning, but the paper highlights unresolved challenges in consensus efficiency, model security, and framework scalability.
- Federated learning background: Federated learning trains models across distributed clients while keeping their data samples local, but network conditions and client availability are unreliable.Applications include video analysis, information inspection, classification, and credit-card fraud detection.
- Decentralized learning: Decentralized approaches seek to reduce or remove the central server by using client-to-client communication or blockchain.Blockchain can store historical operations in a tamper-resistant distributed ledger, while segmented gossip uses node-to-node bandwidth.
- Blockchain-based FL: Blockchain-based federated-learning methods record node performance or updates to support consensus and reduce malicious attacks.Prior proposals include blockchain-maintained global models, parallel learning for multiple models, and consensus based on data amount and historical performance.
- Open challenges: The paper identifies three remaining challenges: consensus efficiency, model security, and framework scalability.Broadcasting consensus is costly with many learning nodes; systems must protect models from unauthorized access and poisoning and adapt to different application scenarios.
III. THE PROPOSED FRAMEWORK
BFLC replaces the centralized federated-learning server with blockchain and smart contracts, while committee consensus handles update selection and block generation. Training nodes locally update the model, submit updates, and receive committee validation before blockchain recording.
- Decentralized architecture: BFLC uses blockchain and smart contracts to implement centralized federated-learning functions without a centralized server.The blockchain maintains global models and local updates while transactions actuate smart-contract operations.
- Training process: Training nodes obtain the newest global model, perform local training, and send local updates to a committee.The committee validates the updates and records a new model or updates on the blockchain.
- Committee consensus: A delegated consensus mechanism addresses gradient selection and block generation to reduce federated-learning communication costs.The framework assigns committee nodes responsibility for consensus-related operations.
A. Blockchain Storage
BFLC uses an alliance blockchain with separate model and update blocks to control access and organize federated-learning information. Smart contracts aggregate verified updates after enough update blocks arrive, while historical blocks support recovery and verification.
- Storage design: BFLC stores learning information on an alliance blockchain accessible only to authorized devices.The system uses separate block types for global models and local updates.
- Block lifecycle: The initial model occupies block #0, nodes train locally, and verified gradients are placed into update blocks.When enough update blocks accumulate, a smart contract triggers aggregation and places the next-round model on the chain.
- Block lifecycle: For round t, block #t×(k +1) stores the model, while blocks #[t×(k +1)+1, (t+1)× (k +1)−1] store that round’s updates.Here, k denotes the required updates per round and t denotes the round number.
- Block contents: An update block contains the round number, local update gradient, uploader address, and update score.The block also includes block headers.
B. Committee Consensus Mechanism
The Committee Consensus Mechanism (CCM) validates local gradients through a small committee before blockchain inclusion, reducing consensus participation while supporting security and flexible training participation.
- CCM validates local gradients before appending them to the blockchain, with a small committee responsible for verification and block generation.Other nodes train locally and send updates to the committee, which scores updates and packs only qualified ones onto the chain.
- Only a few nodes validate updates instead of broadcasting them to every node for agreement.
- Rotating committees provide changing validation sets, yielding a k-fold cross-validation setting in federated learning.Committee members do not participate in local training during the round, so their local data serve as validation data.
- The mechanism addresses non-IID data by validating updates against committee members’ data distributions, while nodes can train and submit updates asynchronously.Only a certain number of updates is required per round, and nodes can obtain the current model and train when available.
- Validated updates are aggregated by the committee into a new global model after enough valid updates are collected, then the model is packed on-chain.
IV. DISCUSSION
BFLC manages participation through verified membership and blockchain-based model access, while using contribution-based rewards to encourage update submission.
- Initial managers verify devices before they join the training community and control participant permissions.
- The latest global model can be quickly found on-chain, and nodes may use it locally or submit updates for committee validation.Aggregation requires only a certain number of valid updates, and only part of the nodes need to be online.
- BFLC proposes profit sharing by contribution to incentivize nodes to provide updates.
- Permission fees provide access to the latest models, while managers distribute post-aggregation rewards according to submitted-update scores.
- Frequent update submission can earn more rewards and attract participation, although fee, sharing, and dividend configurations remain open for study.
B. Committee Election
BFLC elects each new committee from providers of validated updates, and the paper contrasts random, score-based, and multi-factor election strategies by their security, generalization, and overhead trade-offs.
- Each round elects a new committee from providers of validated updates, making election important because the committee determines which updates are aggregated.
- Random election: Random election can improve generalization and reduce overfitting, but offers weak resistance when malicious nodes disguise themselves as normal.
- Election by score: Score-based election increases attack cost and security, but may worsen sample imbalance by excluding some nodes from committees.
- Multi-factor optimization: Multi-factor optimization combines device factors such as transmission rate with validation scores, at the cost of additional computing overhead.Its use should depend on the realistic scenario and associated requirements.
C. Malicious Nodes
The paper analyzes malicious-update acceptance under committee validation and finds that successful committee capture requires a malicious majority, while historical blockchain state permits failback after attacks.
- Under CCM, committee verification occurs before malicious updates are aggregated, enabling theoretical analysis of attack factors and success probability.
- A malicious update is accepted only when more than half of the committee members cooperate with it.
- If the first committee contains more than M^2 honest nodes, malicious nodes cannot enter the committee and harm the global model.
- With similar node performance, attack success is modeled as selecting A × p nodes containing more than half of the A × q malicious nodes.
- Malicious attack probability becomes markedly positive only when the malicious percentage exceeds 50%, requiring 51% of computational resources; historical models and updates support failback.
D. Storage Optimization
BFLC reduces storage demands by allowing capacity-constrained nodes to delete historical blocks while retaining current models and updates. The paper also considers trusted external storage as an alternative, but deletion introduces a credibility trade-off.
- Storage reduction: Nodes with insufficient capacity can delete historical blocks locally and retain only the latest model and current-round updates.Core nodes preserve historical data for disaster recovery and block verification.
- Trade-off: Deleting historical blocks solves local storage shortages but decreases blockchain credibility and may be rejected by mutually distrusting nodes.The scheme preserves recovery and verification capabilities on core nodes rather than on every node.
- External storage: Trusted third-party storage can hold model files and updates while the blockchain records their network addresses and modification operations.The external storage provides disaster-recovery backup and distributed file-storage services.
V. EXPERIMENTAL
Experiments evaluate BFLC on FEMNIST under normal training and malicious-node attacks. BFLC approaches basic FL performance as active participation increases and resists higher malicious-node proportions than the compared methods.
- A. Settings and Normal Training: BFLC was evaluated on FEMNIST, using 900 simulated devices with unbalanced and non-independent local data distributions.The dataset contains 80,5263 samples, 3550 users, and 62 handwritten-character classes.
- A. Settings and Normal Training: As the proportion of active nodes increases, BFLC approaches basic FL accuracy and incurs only a slight loss relative to stand-alone training.The comparison is reported in Table I for different active-node proportions.
- A. Settings and Normal Training: BFLC reduces per-round consensus computation to P × Q when P training nodes use a committee of size Q.The passage contrasts this committee-based calculation with broader consensus communication.
- B. Under Malicious Attack: Under random Gaussian perturbation attacks, BFLC, basic FL, and CwMed were compared across different malicious proportions among active nodes.Basic FL applies no defense, CwMed uses coordinate-wise medians, and BFLC scores updates through committee median local prediction accuracy.
- B. Under Malicious Attack: BFLC resists a much higher malicious-node proportion than the compared methods, even when malicious committee members collude by assigning 90%–100% scores to malicious updates.The active-node proportion is fixed at 10%, with 20% elected as the next-round committee.
VI. CONCLUSION
The conclusion presents BFLC as a decentralized blockchain-based federated-learning framework using committee consensus to address malicious nodes and servers. Experiments support effectiveness on real-world data, while the paper discusses scalability across security, storage, and incentives.
- VI. CONCLUSION: BFLC uses blockchain and committee consensus to provide a decentralized federated-learning framework addressing malicious training nodes and central servers.The framework targets both global-model damage and privacy leakage by malicious participants.
- VI. CONCLUSION: Experiments on a real-world dataset produced a global model similar to centralized training in a federated-learning framework.This is presented as verification of BFLC’s effectiveness.
- VI. CONCLUSION: The paper discusses BFLC scalability through security, data storage, and incentive mechanisms.These areas are identified as having broad research prospects.