Source-linked AI summary
Reliable Federated Learning for Mobile Networks
Jiawen Kang, Zehui Xiong, Dusit Niyato, Yuze Zou, Yang Zhang, Mohsen Guizani
TL;DR
Federated learning needs reliable worker selection because mobile devices can submit malicious or low-quality updates. This paper combines reputation-based selection with consortium-blockchain management and reports more reliable federated learning in mobile networks.
Problem
Federated-learning tasks need to identify trusted workers because mobile devices may submit malicious or unreliable updates.
Method
The paper selects workers using reputation calculated with multi-weight subjective logic and manages reputation through a decentralized consortium blockchain.
Results
Numerical results show that the proposed schemes improve the reliability of federated-learning tasks in mobile networks.
Takeaways & Limitations
Reputation-based selection supports reliable and trusted worker participation in federated learning.
Abstract
from arXiv · showhide
Federated learning, as a promising machine learning approach, has emerged to leverage a distributed personalized dataset from a number of nodes, e.g., mobile devices, to improve performance while simultaneously providing privacy preservation for mobile users. In the federated learning, training data is widely distributed and maintained on the mobile devices as workers. A central aggregator updates a global model by collecting local updates from mobile devices using their local training data to train the global model in each iteration. However, unreliable data may be uploaded by the mobile devices (i.e., workers), leading to frauds in tasks of federated learning. The workers may perform unreliable updates intentionally, e.g., the data poisoning attack, or unintentionally, e.g., low-quality data caused by energy constraints or high-speed mobility. Therefore, finding out trusted and reliable workers in federated learning tasks becomes critical. In this article, the concept of reputation is introduced as a metric. Based on this metric, a reliable worker selection scheme is proposed for federated learning tasks. Consortium blockchain is leveraged as a decentralized approach for achieving efficient reputation management of the workers without repudiation and tampering. By numerical analysis, the proposed approach is demonstrated to improve the reliability of federated learning tasks in mobile networks.
I. INTRODUCTION
Federated learning keeps mobile training data local but remains vulnerable to unreliable updates. The paper proposes reputation-based worker selection with decentralized consortium-blockchain management.
- Motivation: Federated learning lets mobile devices collaboratively train a global model without centrally storing raw training data.Devices train locally, send model updates, and repeat the process with a central aggregator.
- Vulnerabilities: Malicious updates, including poisoning attacks, can mislead the global model and cause collaborative-learning failure.Unreliable behavior may be intentional or unintentional.
- Proposed approach: Reputation is introduced as a metric for selecting reliable and trusted federated-learning workers.Task publishers use workers’ historical behaviors and reputation opinions to identify suitable participants.
- Proposed approach: Consortium blockchain provides decentralized, tamper-resistant, and non-repudiable reputation management.Edge nodes can serve as pre-selected miners using lightweight consensus mechanisms.
- Proposed approach: A multi-weight subjective logic model combines interaction histories with recommended reputation opinions.This provides the basis for calculating each interacting worker’s reputation.
II. FEDERATED LEARNING AND ITS VULNERABILITIES
Federated learning addresses privacy concerns by training models across mobile devices while keeping user data local. The section highlights applications including keyboards, recommendations, traffic prediction, and healthcare.
- Federated learning: Federated learning enables distributed mobile devices to train a global model using locally stored raw data.This avoids centralized collection of sensitive user information.
- Applications: Gboard uses federated learning to improve language-model quality while training input data locally.The application provides security and privacy protection for users.
- Applications: Service recommendation systems can train on searching and location histories without directly exposing users’ private data.Federated learning supports collaborative recommendation-model training across mobile devices.
- Applications: Distributed traffic monitoring and prediction can use vehicle sensing data without direct access to personal vehicle data.The passage describes this as an application of federated learning to privacy-sensitive traffic information.
- Applications: Mobile healthcare can collaboratively train models from local health data without centrally collecting sensitive patient information.Hospitals and researchers can use federated learning for healthcare analytics.
B. Security Challenges and Motivations
Mobile federated learning faces unreliable workers, malicious updates, and inadequate monitoring. Existing approaches lack reliable trust metrics, universal selection methods, and timely behavioral monitoring.
- Security threats: Malicious or unreliable data owners may provide deceptive sensing data or manipulate data during insecure transmission.Selecting such workers can enable poisoning or collusive attacks.
- Assumption: The analysis assumes that the central aggregator is not compromised or malicious.This defines a security boundary for the proposed setting.
- Open challenges: Existing federated-learning systems lack reliable and fair metrics for evaluating workers’ trustworthiness.Random selection based on verifiable random functions or resource conditions cannot remove unreliable workers.
- Open challenges: Existing worker-selection schemes are either centralized or allow devices to join training at will.Both approaches remain exposed to unreliable or untrusted workers in mobile networks.
- Open challenges: Central aggregators struggle to monitor large-scale worker behavior in real time.Without accumulated historical metrics and synchronized malicious-worker information, unreliable workers may be selected again.
III. REPUTATION MANAGEMENT FOR RELIABLE FEDERATED LEARNING
The paper adapts reputation management from crowdsensing to federated learning. Subjective logic and consortium blockchain are used to support reliable worker selection and decentralized reputation handling.
- Background: Prior crowdsensing studies use reputation to identify honest providers and evaluate data quality.Reputation supports selecting providers likely to submit high-quality data.
- Federated-learning adaptation: The paper uses workers’ model-training histories and past reliable or unreliable behavior to construct reputation for federated learning.Task publishers can use this metric to select trusted workers and reduce poisoning or low-quality data contributions.
- Reputation architecture: Subjective logic provides a decentralized reputation-calculation method for avoiding risks of centralized reputation management.The approach is paired with consortium blockchain properties of immutability and decentralization.
- Reputation architecture: Consortium blockchain uses lightweight, fast consensus that is described as practical for mobile networks.The scheme relies on pre-selected miners and supports secure reputation storage.
B. Reputation-based Worker Selection Scheme with Consortium Blockchain
The scheme selects federated-learning workers through reputation-based screening and updates reputation records using a consortium blockchain. It combines task publication, worker selection, federated training, and signed reputation updates.
- Task publication: Task publishers broadcast data requirements, and eligible mobile devices return joining requests containing identity and data-resource information.
- Worker selection: Worker candidates are selected when their reputation values exceed a publisher-defined threshold, which may reflect security requirements or candidate statistics.
- Reputation calculation: Composite reputation combines local opinions from interaction histories with recommended opinions obtained from the consortium reputation blockchain.
- Federated learning: Selected workers receive an initial global SGD model and collaboratively train it using local data and iterative gradient updates.For ε-accuracy, SGD requires O(µ^2/ε) iterations on each worker, where µ is the condition number.
- Reputation updating: Signed interaction histories and local opinions are uploaded as transactions, validated through consensus such as PBFT, and stored in new reputation blocks.
IV. EFFICIENT REPUTATION CALCULATION SCHEME
The reputation calculation scheme forms worker opinions using subjective logic and combines local and recommended information into composite reputations for worker selection. Reputation is represented through belief, distrust, and uncertainty, with the calculation workflow integrating interaction evidence and publisher recommendations.
- Subjective logic: Subjective logic represents a publisher’s view of a worker with belief, distrust, and uncertainty, which are integrated into a composite reputation value.
- Interaction evidence: Training iterations become positive or negative interaction events according to whether a publisher perceives a worker’s local model update as reliable.
- Recommended opinions: Recommended opinions are integrated into an overall recommended reputation opinion before composite reputation is formed.
- Weight integration: The calculation integrates interaction timelines and interaction effects into an overall weight for reputation opinion calculation.
- Calculation workflow: The workflow calculates local reputation opinions, integrates recommended opinions from other publishers, and forms composite reputation for worker selection.
B. Multi-weight Subjective Logic Model
The multi-weight subjective logic model improves reputation calculation by weighting interaction attributes. It considers interaction frequency, timing, and effects to produce more accurate and reliable reputation opinions.
- Model purpose: Multi-weight subjective logic accounts for different interaction-event attributes to calculate more accurate and reliable reputation opinions.
- Interaction Frequency: Interaction frequency measures publisher-worker familiarity relative to the publisher’s average interactions with other workers during a time window.
- Interaction Timelines: Interaction timelines divide events into recent and past interactions because a worker’s trust level and local reputation opinion change over time.
- Interaction Effects: Interaction effects classify events as positive or negative, with negative interactions decreasing worker reputation and positive interactions receiving higher weight.
C. Recommended Reputation Opinions
Recommended reputation opinions from other task publishers are combined into an overall recommended opinion using publisher-specific weights. This opinion is represented by belief, distrust, and uncertainty degrees.
- Other publishers’ local reputation opinions are treated as recommended opinions and integrated according to the receiving publisher’s weights.
- The overall recommended opinion contains recommended belief, distrust, and uncertainty degrees calculated by weighted arithmetic mean.
D. Combining Local Reputation Opinions with Recommended Reputation Opinions
The task publisher combines recommended opinions with its own local reputation opinion to form a composite reputation for each worker. This composite includes belief, distrust, and uncertainty degrees.
- The composite reputation incorporates both overall recommended opinions and the task publisher’s local reputation opinion.This design aims to avoid collusion cheating by other task publishers.
- The resulting reputation is represented as a final reputation opinion vector with belief, distrust, and uncertainty degrees.
V. NUMERICAL RESULTS
Simulations evaluate federated learning with malicious attackers, unreliable data, and reputation-based defenses. Accuracy declines with stronger attacks and poorer data, while reputation thresholds and reputation dynamics affect detection and accuracy.
- Simulation Setting: The simulation uses MNIST with ten workers: two malicious, four unreliable, and four well-behaved.Unreliable workers receive data from limited classes, while malicious workers modify labels to create poisoning attacks.
- Performance Results: 76.12% accuracy with two attackers is 7.7% lower than with one attacker at EMD 1.6 and attack strength 0.9.
- Performance Results: Increasing EMD, attacker number, or attack strength decreases federated learning accuracy.The results attribute negative impacts to both low-quality training data and poisoning attacks.
- Performance Results: Reputation decreases under misbehavior in MSL, TSL, and ATV, whereas it increases linearly without reputation defenses.MSL decreases more sharply than ATV and TSL over a short time because it considers interaction effects, frequency, and timeline.
- Performance Results: Higher reputation thresholds improve accuracy; MSL matches ATV above threshold 0.35, while TSL, MSL, and ATV match above 0.45.At high EMD and attack strength, malicious and unreliable workers are easier to detect and remove.
VI. CONCLUSION AND FUTURE DIRECTIONS
The paper addresses reliable worker selection in mobile-network federated learning through reputation and consortium blockchain. It reports improved reliability while identifying future work on validation, worker-number optimization, and dynamic optimization.
- Conclusion: A reputation-based scheme selects reliable and trusted workers for federated learning in mobile networks.
- Conclusion: Multi-weight subjective logic calculates worker reputations, while consortium blockchain manages them decentrally with tamper resistance and non-repudiation.
- Conclusion: Numerical results show that the proposed schemes bring reliable federated learning to mobile networks.
- Future Directions: Future work includes more accurate non-IID poisoning validation, efficient worker-number optimization, and dynamic optimization.