Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

TPFL: A Trustworthy Personalized Federated Learning Framework via Subjective Logic

Jinqian Chen, Jihua Zhu

TL;DR

This paper introduces Trustworthy Personalized Federated Learning framework designed for classification tasks via subjective logic, which adopts a unique approach by employing subjective logic to construct federated models, providing probabilistic decisions coupled with an assessment of uncertainty rather than mere probability assignments.

Abstract

Federated learning (FL) enables collaborative model training across distributed clients while preserving data privacy. Despite its widespread adoption, most FL approaches focusing solely on privacy protection fall short in scenarios where trustworthiness is crucial, necessitating advancements in secure training, dependable decision-making mechanisms, robustness on corruptions, and enhanced performance with Non-IID data. To bridge this gap, we introduce Trustworthy Personalized Federated Learning (TPFL) framework designed for classification tasks via subjective logic in this paper. Specifically, TPFL adopts a unique approach by employing subjective logic to construct federated models, providing probabilistic decisions coupled with an assessment of uncertainty rather than mere probability assignments. By incorporating a trainable heterogeneity prior to the local training phase, TPFL effectively mitigates the adverse effects of data heterogeneity. Model uncertainty and instance uncertainty are further utilized to ensure the safety and reliability of the training and inference stages. Through extensive experiments on widely recognized federated learning benchmarks, we demonstrate that TPFL not only achieves competitive performance compared with advanced methods but also exhibits resilience against prevalent malicious attacks, robustness on domain shifts, and reliability in high-stake scenarios.

TPFL: A Trustworthy Personalized Federated Learning Framework via Subjective Logic

TL;DR

This paper introduces Trustworthy Personalized Federated Learning framework designed for classification tasks via subjective logic, which adopts a unique approach by employing subjective logic to construct federated models, providing probabilistic decisions coupled with an assessment of uncertainty rather than mere probability assignments.

Abstract

Federated learning (FL) enables collaborative model training across distributed clients while preserving data privacy. Despite its widespread adoption, most FL approaches focusing solely on privacy protection fall short in scenarios where trustworthiness is crucial, necessitating advancements in secure training, dependable decision-making mechanisms, robustness on corruptions, and enhanced performance with Non-IID data. To bridge this gap, we introduce Trustworthy Personalized Federated Learning (TPFL) framework designed for classification tasks via subjective logic in this paper. Specifically, TPFL adopts a unique approach by employing subjective logic to construct federated models, providing probabilistic decisions coupled with an assessment of uncertainty rather than mere probability assignments. By incorporating a trainable heterogeneity prior to the local training phase, TPFL effectively mitigates the adverse effects of data heterogeneity. Model uncertainty and instance uncertainty are further utilized to ensure the safety and reliability of the training and inference stages. Through extensive experiments on widely recognized federated learning benchmarks, we demonstrate that TPFL not only achieves competitive performance compared with advanced methods but also exhibits resilience against prevalent malicious attacks, robustness on domain shifts, and reliability in high-stake scenarios.

Paper Structure

This paper contains 25 sections, 1 theorem, 17 equations, 7 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Trustworthy Personalized Federated Learning via Subjective Logic
  4. Forming Opinions via Subjective Logic
  5. Robust Local Training with Trainable Heterogeneity Prior
  6. Debiased Trustworthy Inference via Instance Uncertainty-Informed Opinion Fusion
  7. Secure Aggregation with Model Uncertainty-based Anomaly Detection
  8. Experiments
  9. Predictive Performance
  10. Reliability
  11. Security
  12. Additional Cost
  13. Ablation Study
  14. Conclusion
  15. Detailed Related Work
  16. ...and 10 more sections

Key Result

Proposition 1

The evidence of aggregated opinion calculated through instance uncertainty-informed opinion fusion is equivalent to the confidence-based weighted average on observed evidence, that is:

Figures (7)

  • Figure 1: Difference Between Subjective Models and Traditional Models. Traditional models directly estimate the first-order probability distribution through Softmax operation. In contrast, subjective models observe evidence from the instance to form opinions, constructing a Dirichlet distribution (i.e., second-order distribution) on class assignments to allow the estimation of uncertainty.
  • Figure 2: Overall framework of Trustworthy Personalized Federated Learning (TPFL). In the local training stage (red part), data is directly utilized to train the subjective model and adjust its local prior distribution (See section \ref{['sec3.2']} for detailed loss designs). The local model is further uploaded to the central server for aggregation (green part). To ensure training safety, TPFL utilizes an out-of-distribution (OOD) dataset with a small size to evaluate each uploaded model. Leveraging the sensitive property of subjective models (See section \ref{['sec3.2']}), evidences are first filtered to exclude updates with extremely large or overflow values, protecting the aggregated model from attacks utilizing significant tempering ModelReplacement. The remaining opinions are further integrated via instance uncertainty-guided trustworthy fusion. Discrepancies between original and fused opinions are treated as model uncertainty to measure the similarity of updates. Too similar updates with small model uncertainty deviation will be filtered to counter accumulation attacks LIELocalModelPoisoningMPAF. Finishing all training, the local models will be required to fine-tune on re-balanced datasets, and the inference (blue part) will integrate three sources to produce reliable predictions.
  • Figure 3: Reliability experimental results. Fig.\ref{['ra']} and \ref{['rb']} demonstrate the effectiveness of estimated uncertainty. With the decrease of the threshold, the predictive accuracy significantly increases. Fig.\ref{['rc']} and \ref{['rd']} validate the reliability of TPFL on OOD datasets, showcasing higher uncertainty compared with other methods.
  • Figure 4: Safety performance of TPFL. We evaluate the performance under different attack strategies with varied malicious client ratios from 0.1 to 0.5 and compare the performance with SOTA defense methods. As can be seen, TPFL exhibits great resistance against SOTA attacking strategies, showcasing similar safety performance with advanced defense methods.
  • Figure 5: Ablation study on the sensitivity of hyper-parameters. As can be seen, TPFL is robust to the choice of $\lambda_2$ and $\lambda_3$, while $\lambda_1$ needs a proper choice to balance suppressing fake evidence and encouraging true evidence (e.g., 0.1).
  • ...and 2 more figures

Theorems & Definitions (5)

  • Definition 1: Multinomial Opinion of Subjective Federated Model
  • Definition 2: Bijective Mapping Between Multinomial Opinion and Dirichlet PDF
  • Definition 3
  • Definition 4
  • Proposition 1