Research map / Federated learning and privacy / Personalized and heterogeneous FL
Fine-Tuning Personalization in Federated Learning to Mitigate Adversarial Clients
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Analyzes interpolated personalized federated learning with Byzantine clients, deriving when full collaboration is worse than fine-tuned collaboration level, and validates on mean estimation, Phishing and MNIST.
Problem
Personalized federated learning where some clients are adversarial (Byzantine): full collaboration can perform worse than local learning because adversaries induce optimization error that grows with data heterogeneity, so the right collaboration level must be chosen.
Approach
Interpolate each client's local loss with the average loss of correct clients, using a robust aggregator for the collaborative part; derive optimization and generalization bounds showing the optimal collaboration level shrinks with heterogeneity and with the adversarial fraction (via a kappa-robustness factor), and tune lambda accordingly.
Key results
- On Phishing and MNIST with n=20 clients, intermediate collaboration levels lambda strictly between 0 and 1 beat both full collaboration and local learning for moderate adversary counts (f=3 or 6).
- With f=9 adversaries, robust federated learning falls below 50% local test accuracy, while tuning the collaboration degree gives substantial gains over robust FL.
- In mean estimation, choosing the right collaboration level reduces error by 50% compared to the local estimator at f=150 adversaries.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| choose the collaboration level instead of using local estimation (λ=0) | Robust mean estimation with sign-flipping adversaries · Synthetic Gaussian mean estimation, f=150 | mean squared error: reduces the error by 50% | Helped |
| use full collaboration (λ=1) rather than a lower collaboration level | Robust mean estimation · Low heterogeneity | mean squared error | No clear effect |
| use full collaboration rather than fine-tuned collaboration | Personalized federated learning with adversarial clients · Phishing and MNIST local test datasets | accuracy | Hurt |
| use robust federated learning with full collaboration rather than local learning | Robust federated learning · Phishing and MNIST, f=9 | accuracy: under 50% | Hurt |
| reduce the collaboration level from full collaboration | Personalized federated learning with adversarial clients · Phishing and MNIST, f=9 | accuracy | Helped |
| use an intermediate collaboration level (0<λ<1) rather than local learning or full collaboration | Personalized federated learning with adversarial clients · Phishing and MNIST, data scarcity, f=3 or f=6 | accuracy | Helped |
| increase the collaboration level as task complexity increases | Robust mean estimation · Synthetic Gaussian mean estimation | mean squared error | Mixed |
| increase the adversarial-client fraction while holding other settings fixed | Robust mean estimation · Synthetic Gaussian mean estimation | mean squared error | Mixed |
| increase data heterogeneity | Personalized federated learning with adversarial clients · Phishing and MNIST local test datasets | accuracy | Hurt |
Surprising findings
- Even when heterogeneity is low, full collaboration is not optimal under adversaries, and local learning can beat state-of-the-art robust FL when the local task is simple.
Limitations
- The interpolated algorithm is not communication-efficient and may not scale to high-dimensional tasks with many clients.
- Analysis covers only the interpolation personalization strategy, not regularization, clustering, or layer-specialization methods.
- Experiments are small-scale (n=20 clients, 5 runs, simple models).
Experimental setup
Datasets: Phishing MNIST synthetic Gaussian mean estimation
Benchmarks: Phishing MNIST synthetic Gaussian mean estimation
Metrics: squared error of mean estimate local test accuracy
Baselines: local learning full collaboration robust FL NNM + trimmed mean
Builds on: robust distributed gradient descent NNM pre-aggregation domain adaptation discrepancy theory sign-flipping attack
Closest papers by meaning
- Bridging Generalization Gap of Heterogeneous Federated Clients Using Generative Models (ICLR 2026 · similarity 0.91)
- Personalized Federated Learning with Mixture of Models for Adaptive Prediction and Model Fine-Tuning (NeurIPS 2024 · similarity 0.89)
- Personalized Federated Learning via Feature Distribution Adaptation (NeurIPS 2024 · similarity 0.88)
- Flexible Sharpness-Aware Personalized Federated Learning (AAAI 2025 · similarity 0.88)
- FedAS: Bridging Inconsistency in Personalized Federated Learning (CVPR 2024 · similarity 0.88)
- Understanding the Statistical Accuracy-Communication Trade-off in Personalized Federated Learning with Minimax Guarantees (ICML 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.