PaperlayerView exact revision on arXiv

Source-linked AI summary

Reliable Federated Learning for Mobile Networks

Jiawen Kang, Zehui Xiong, Dusit Niyato, Yuze Zou, Yang Zhang, Mohsen Guizani

arXiv:1910.06837v1cs.CR

TL;DR

Mobile federated learning needs ways to identify trusted workers, but central aggregators lack timely monitoring and accumulated behavioral evidence. The paper uses reputation-based selection and consortium blockchain for reputation management, and reports that its MSL scheme calculates reputation more accurately and fairly, leading to more reliable worker selection.

  • Problem

    Mobile federated learning lacks timely, dynamic worker monitoring and accumulated behavioral metrics for identifying unreliable workers and selecting trusted ones.

    Source: B. Security Challenges and Motivations · p. 3

  • Method

    The paper selects workers by reputation and uses a consortium blockchain at edge nodes to manage reputations in a decentralized manner.

    Source: §I · pp. 1–2

  • Results

    The MSL scheme is reported to achieve more accurate and fair reputation calculation, leading to more reliable worker selection in federated learning.

    Source: B. Performance Results · p. 7

  • Takeaways & Limitations

    The authors conclude that their schemes can bring reliable federated learning to mobile networks.

    Source: §VI · p. 7

Abstract

from arXiv · show

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

The article proposes reputation-based worker selection to reduce unreliable updates in federated learning, with subjective-logic reputation opinions and a consortium blockchain for decentralized, tamper-resistant reputation management.

I. INTRODUCTION

  • Federated learning avoids centrally storing raw training data by having devices train locally and send model updates to an aggregator, reducing privacy-leakage risk.The aggregator combines updates into a new global model, and the process repeats until the model reaches a certain accuracy.Source: §I · p. 1
  • Federated learning remains vulnerable to intentional poisoning updates and unintentional low-quality updates caused by high-speed mobility or energy constraints.Both kinds of unreliable local updates can adversely affect the global model or collaborative learning.Source: §I · p. 1
  • Reputation-based selection lets task publishers choose high-reputation workers to limit unreliable updates and improve learning-task accuracy.Publishers calculate reputation opinions using their own interaction histories and opinions recommended by other publishers.Source: §I · pp. 1–2
  • A consortium blockchain deployed at edge nodes records and manages worker reputations in a decentralized, non-repudiable, tamper-resistant ledger.Edge nodes such as roadside units and base stations serve as pre-selected miners with storage and computation resources.Source: §I · pp. 1–2
  • A multi-weight subjective logic model calculates reputation from publishers’ interaction histories and recommended reputation opinions.The model integrates each publisher’s own past-interaction opinions with recommendations from other task publishers.Source: §I · pp. 1–2

II. FEDERATED LEARNING AND ITS VULNERABILITIES · A. Federated Learning and Its Mobile Applications

Federated learning lets mobile devices train a shared model by sending local updates while retaining raw data locally, addressing privacy and data-collection challenges associated with centralized machine learning. The section illustrates its use in keyboard, recommendation, traffic-prediction, and healthcare applications.

A. Federated Learning and Its Mobile Applications

  • Mobile devices train a global model by keeping raw data locally, sending local updates, and having a central aggregator average those updates.The local updates may be computed using distributed Stochastic Gradient Descent; this approach blocks direct access to raw training data.Source: A. Federated Learning and Its Mobile… · p. 2
  • Centralized machine learning faces single points of failure, sensitive-data leakage, and substantial costs for collecting and storing training data.Source: A. Federated Learning and Its Mobile… · p. 2
  • Google's Gboard uses federated learning to improve language-model quality while training input data locally for user security and privacy.Source: A. Federated Learning and Its Mobile… · p. 2
  • Federated learning lets mobile devices train entertainment and restaurant recommendation models without exposing users' search and location histories.Source: A. Federated Learning and Its Mobile… · p. 2
  • For traffic prediction, federated learning can use vehicle data without direct access to personal information, protecting privacy while enhancing prediction accuracy.UberEATs uses real-time traffic information to estimate food-delivery times.Source: A. Federated Learning and Its Mobile… · p. 2
  • In mobile healthcare, federated learning supports collaborative model training on local health data without central collection of sensitive patient records.NVIDIA Clara is used to deploy tasks recommending treatments or determining biomarkers automatically.Source: A. Federated Learning and Its Mobile… · p. 2

B. Security Challenges and Motivations

Unreliable workers can harm federated-model quality through both malicious and inadvertent updates, motivating reliable worker selection. The section identifies missing trust metrics, universal selection schemes, and timely worker monitoring as key challenges.

Security challenges and motivations

  • Malicious sensing or tampered devices and insecure channels, can enable poisoning,, while energy limits or mobility cause low-quality updates that threaten global-model quality.Poisoning can involve tampering with part of the training data or injecting poisonous data to increase misclassification and manipulate training results.Source: B. Security Challenges and Motivations · p. 2

No reliable and fair metrics to evaluate workers

No efficient and universal worker selection schemes

No timely monitoring methods for workers

  • Without real-time monitoring and historical performance metrics, malicious or unreliable workers may be selected again; the section proposes a reliable metric and worker-selection scheme.The challenge also includes lacking synchronous information about malicious and unreliable worker lists.Source: B. Security Challenges and Motivations · p. 3

III. REPUTATION MANAGEMENT FOR RELIABLE FEDERATED LEARNING · A. Overview of Reputation Management in Crowdsensing · B. Reputation-based Worker Selection Scheme with Consortium Blockchain

The paper adapts reputation-based provider selection from crowdsensing to federated learning, using a consortium blockchain for decentralized reputation management. Its workflow selects workers by reputation, filters unreliable updates during training, and records interaction histories to inform future selection.

A. Overview of Reputation Management in Crowdsensing

  • Crowdsensing studies use reputation to assess data-provider honesty and quality, with schemes based on credit and trajectory matching, contribution history, or collaborative scoring,,,,.These approaches aim to select higher-quality providers, encourage skilled workers, or quantify data trustworthiness.Source: A. Overview of Reputation Management… · p. 3
  • The paper applies reputation to federated-learning worker selection and combines subjective logic with a decentralized consortium blockchain to manage reputation securely and efficiently,,.The blockchain is described as immutable and decentralized, with pre-selected miners enabling lightweight, faster consensus for mobile networks.Source: A. Overview of Reputation Management… · pp. 3–4

B. Reputation-based Worker Selection Scheme with Consortium Blockchain

  • Task publishers broadcast data requirements, validate applicants’ identities and resources, then select eligible worker candidates whose reputation exceeds a publisher-set threshold.Thresholds may differ by publisher or be adjusted using the candidates’ mean and standard deviation of reputation values.Source: B. Reputation-based Worker Selection… · p. 4
  • Each candidate’s final reputation combines a local opinion derived from interaction history with recommended opinions downloaded from the latest reputation-blockchain data.The subjective logic model uses three weights about past interactions to form local opinions.Source: B. Reputation-based Worker Selection… · p. 4
  • Selected workers collaboratively train a global SGD model on local data; publishers validate computation time, detect lazy or poisoning updates, average remaining updates, and record negative interactions.The workflow identifies RONI for IID scenarios and FoolsGold for non-IID scenarios; the paper assumes sufficiently uniform device data distribution for RONI gradient validation.Source: B. Reputation-based Worker Selection… · p. 4
  • Publishers upload digitally signed interaction histories and local reputation opinions to pre-selected miners, who use consensus such as PBFT to add them to the blockchain for future worker selection.All task publishers can then obtain the latest reputation opinions for a worker candidate.Source: B. Reputation-based Worker Selection… · p. 5

IV. EFFICIENT REPUTATION CALCULATION SCHEME

The scheme calculates composite worker reputations from task publishers’ opinions to support secure worker selection. It uses subjective logic and stores reputation opinions on a decentralized blockchain.

IV. EFFICIENT REPUTATION CALCULATION SCHEME

  • Task publishers’ reputation opinions are collected and integrated into composite reputation values for secure worker selection.Each publisher calculates worker values using its local reputation opinion and recommended opinions.Source: §IV · p. 5
  • Subjective logic, used to evaluate trust between network entities,, represents opinions as belief, distrust, and uncertainty.A publisher’s belief in a vehicle rises when its updates are high-quality without unstable communication impacts, and vice versa.Source: §IV · p. 5
  • All task publishers’ reputation opinions are securely updated and stored in a decentralized reputation blockchain.Source: §IV · p. 5

A. Subjective Logic Model for Reputation Calculation

The model derives workers’ local reputation opinions from task publishers’ assessments of update reliability and communication quality. Each opinion combines belief, distrust, and uncertainty into a local reputation value.

A. Subjective Logic Model for Reputation Calculation

  • Publishers classify each worker update as a positive or negative interaction using poisoning-attack detection and proof of elapsed time, then record counts α_j and β_j.The resulting counts are used to generate local reputation opinions for each worker.Source: A. Subjective Logic Model for… · p. 5
  • Each local opinion contains belief, distrust, and uncertainty degrees whose sum is one.The publisher records these opinions after a learning task.Source: A. Subjective Logic Model for… · p. 5
  • Following,, uncertainty reflects the unsuccessful probability of packet transmission between a worker and publisher.Belief and distrust are the positive and negative interaction percentages among interactions with good communication quality: b_i→j = (1 − u_i→j) α_j/(α_j + β_j) and d_i→j = (1 − u_i→j) β_j/(α_j + β_j).Source: A. Subjective Logic Model for… · p. 5
  • The local reputation value is T_i→j = b_i→j + γu_i→j, where γ specifies the effect of uncertainty on reputation.It represents the publisher’s expected belief that the worker provides high-quality local model updates.Source: A. Subjective Logic Model for… · p. 5

B. Multi-weight Subjective Logic Model

The multi-weight subjective logic model extends subjective logic by using interaction frequency, timing, and effects to calculate worker reputation opinions. Recent and positive interactions carry greater weight, while negative interactions reduce reputation.

B. Multi-weight Subjective Logic Model

  • Multi-weight subjective logic extends subjective logic by considering different attributes of interaction events as weights when calculating reputation opinions, aiming to make reputation calculations more accurate and reliable.Fig. 2 shows these three attributes as weights for calculating reputation opinions.Source: B. Multi-weight Subjective Logic Model · p. 5
  • Interaction frequency measures a publisher’s interactions with a worker relative to its average interactions with other workers during a time window.Higher frequency indicates greater familiarity and prior knowledge, leading to a higher local reputation opinion.Source: B. Multi-weight Subjective Logic Model · p. 5
  • A time scale separates recent from past interactions, with recent interactions receiving greater weight in the publisher’s reputation opinion.The passage gives three days as an example time scale.Source: B. Multi-weight Subjective Logic Model · p. 5
  • Negative interactions decrease worker reputation, while positive interactions receive greater weight in reputation-opinion calculation.Interaction timelines and effects upgrade interaction frequency, which is then used to generate an overall weight for local and recommended reputation calculations.Source: B. Multi-weight Subjective Logic Model · pp. 5–6

C. Recommended Reputation Opinions … A. Simulation Setting

The method forms recommended reputation opinions by weighted averaging publishers’ local opinions, then combines them with a publisher’s own opinion to select trusted workers. The supplied passages describe an MNIST simulation setup but provide no numerical results.

C. Recommended Reputation Opinions

  • Recommended belief, distrust, and uncertainty degrees are weighted arithmetic means of other task publishers’ corresponding local opinion degrees.The publisher weights each recommended opinion when forming the overall recommendation.Source: C. Recommended Reputation Opinions · p. 6

D. Combining Local Reputation Opinions with Recommended Reputation Opinions

  • Combining recommended opinions with the publisher’s local opinion yields a final belief-distrust-uncertainty vector, whose belief and uncertainty determine composite reputation.The scheme selects high-reputation workers to defend against intentional or unintentional unreliable local updates; calculation details cite,.Source: D. Combining Local Reputation Opinions… · p. 6

A. Simulation Setting

  • The evaluation simulates digit classification on MNIST, with 60,000 training examples and 10,000 test examples.The simulations use TensorFlow 1.12.0.Source: A. Simulation Setting · p. 6
  • The task has two poisoning attackers, four low-quality-data workers, and four well-behaved workers; attack strength is the percentage of intentionally modified labels.Well-behaved workers’ training data is uniformly distributed across 10 classes, while attackers receive 10-class data before modifying labels; unreliable workers receive randomly assigned subsets of classes, with quality measured by EMD against the population distribution [14].Source: A. Simulation Setting · p. 6
  • Workers produce local SGD updates from random batches of 32 examples, and each global model uses five synchronous iterations,.Local training computation overhead is treated as constant; the reputation blockchain uses Hyperledger Fabric v1.4.0 with PBFT consensus [1],.Source: A. Simulation Setting · p. 6
  • The reputation simulation assumes 20–40 weekly interactions per publisher-worker pair, packet transmission failure probabilities from 0% to 40%, and initial worker reputations of 0.5.It compares Multi-weight Subjective Logic using parameters from with Traditional Subjective Logic and Aggregated Trust Value [15], which aggregates weighted trust-value offsets based on positive and negative events.Source: A. Simulation Setting · p. 6

B. Performance Results

Higher reputation thresholds improve federated learning accuracy, and MSL, TSL, and ATV perform equally above 0.45. The results also show that increasing EMD, attacker count, or attack strength reduces accuracy, while MSL reputation decreases more sharply in the short term.

B. 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.Accuracy decreases as EMD, attacker number, or attack strength increases; unreliable workers with low-quality training data harm learning accuracy.Source: B. Performance Results · p. 7
  • MSL, TSL, and ATV reputations decrease when a worker misbehaves, while reputation without defenses continues to increase.The worker first performs well in six tasks, then trains on poisoning or unreliable examples for 30 task publishers with probability 0.8. MSL drops more sharply in the short term, but ATV drops faster after 12 iterations because it focuses only on interaction effects when calculating trust offsets.Source: B. Performance Results · p. 7
  • For federated learning (EMD=1.6, attack strength=0.9), higher reputation thresholds increase accuracy; MSL is below ATV at lower thresholds, matches ATV above 0.35, and TSL, MSL, and ATV perform equally above 0.45.The text attributes ATV’s false positives to ignoring good workers’ well-behaved histories.Source: B. Performance Results · p. 7

VI. CONCLUSION AND FUTURE DIRECTIONS · BIOGRAPHIES

The article proposes reputation-based worker selection, using multi-weight subjective logic and a consortium blockchain for decentralized reputation management. Numerical results indicate that the schemes can bring reliable federated learning to mobile networks; future work targets validation, worker-count optimization, and reputation thresholds.

VI. CONCLUSION AND FUTURE DIRECTIONS

  • Numerical results showed that the proposed schemes can bring reliable federated learning to mobile networks.The conclusion provides no specific quantitative comparison or metric values.Source: §VI · p. 7
  • The article addresses reliable federated learning by proposing a reputation-based scheme to select trusted workers.The scheme focuses on worker selection in mobile networks.Source: §VI · p. 7
  • Workers’ reputations are calculated using a multi-weight subjective logic model.Source: §VI · p. 7
  • A consortium blockchain manages reputation decentrally with tamper resistance and non-repudiation.The article presents this approach as supporting efficient and secure reputation management.Source: §VI · p. 7
  • Future work should develop more accurate and efficient non-IID validation schemes to address RONI’s model-update validation limitation and improve poisoning-attack detection.Source: §VI · p. 7
  • Future work should optimize the number of participating workers to balance learning performance against resource cost under high worker overhead.Source: §VI · p. 7
  • Future work should dynamically optimize the reputation threshold to minimize malicious workers’ negative effects, potentially using advanced machine learning methods.Source: §VI · p. 7

Related papers

Loading 1910.06837v1…