Research map / Federated learning and privacy / Personalized and heterogeneous FL
pFedMxF: Personalized Federated Class-Incremental Learning with Mixture of Frequency Aggregation
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.
Proposes pFedMxF, a federated class-incremental learning method that assigns randomly sampled, non-overlapping 2D-DFT frequency coefficients to client-task pairs for interference-free PEFT updates, plus an Auto-task Agnostic Classifier for routing.
Problem
Federated class-incremental learning must handle non-IID clients (spatial heterogeneity), sequential tasks (temporal heterogeneity), and heterogeneous client resources, while avoiding interference, catastrophic forgetting, and high communication cost.
Approach
Represent each client-task parameter update as a sparse set of 2D DFT coefficients sampled without replacement, so updates occupy orthogonal frequency subspaces; aggregation is a lossless mixture of these frequency components. A task-routing classifier (AAC) with per-task heads and prototype-based routing handles class-incremental prediction.
Key results
- On CIFAR-100 under quantity-based label imbalance (alpha=6), pFedMxF reaches 71.3% last and 80.7% average accuracy, versus 70.5% last for InfLoRA and 69.3% for PILoRA.
- At the most severe QBLI setting (alpha=2) on CIFAR-100, pFedMxF gets 64.9% average accuracy, while L2P collapses to 5.6%.
- Ablation: pFedMxF with 3000 frequencies (3.2 per-mille sparsity) matches LoRA rank r=4 performance with about half the parameters.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use pFedMxF instead of the strongest reported baseline | CIFAR-100 · QBLI α=6, α=4, and α=2 | Avg. accuracy (%): 78.5 -> 80.7; 75.6 -> 76.2; 62.8 -> 64.9 | Helped |
| use pFedMxF instead of the strongest reported baseline | CIFAR-100 · DBLI β=0.5, β=0.1, and β=0.05 | Avg. accuracy (%): 78.5 -> 80.3; 73.8 -> 75.2; 69.3 -> 70.5 | Helped |
| use pFedMxF instead of the strongest reported baseline | TinyImageNet · QBLI α=6, α=4, and α=2 | Avg. accuracy (%): 81.5 -> 82.7; 81.4 -> 83.1; 77.6 -> 78.4 | Helped |
| use pFedMxF instead of the strongest reported baseline | TinyImageNet · DBLI β=0.5, β=0.1, and β=0.05 | Avg. accuracy (%): 81.6 -> 82.4; 80.1 -> 82.3; 80.3 -> 81.9 | Helped |
| add AAC instead of using a unified classifier (ablation) | pFedMxF · CIFAR-100, QBLI α=6 | Avg. accuracy (%): 74.1 -> 80.7 | Helped |
| assign non-shared random rather than shared frequency coordinates (ablation) | pFedMxF · CIFAR-100, QBLI α=6 | Avg. accuracy (%): 78.2 -> 80.7 | Helped |
| replace FedAvg or FedStack aggregation with pFedMxF aggregation (ablation) | CIFAR-100 · QBLI α=6 | Avg. accuracy (%): 77.8 -> 80.7; 78.6 -> 80.7 | Helped |
| vary the number of frequency components from 1000 to 3000 to 6000 (ablation) | pFedMxF · CIFAR-100, QBLI α=6 | Avg. accuracy (%): 80.2; 80.7; 80.6 | Mixed |
Surprising findings
- L2P collapses to 2.6% last-task accuracy on CIFAR-100 under severe quantity-based imbalance (alpha=2).
- Sharing frequency coordinates across clients (breaking orthogonality) degrades accuracy notably, especially under severe heterogeneity.
Limitations
- Experiments use only simulated non-IID partitions of three image datasets with a single ViT-B/16 backbone.
- Gains over LoRA-based baselines such as InfLoRA and PILoRA are modest in several settings, and the paper reports no statistical significance tests beyond three seeds.
Experimental setup
Datasets: CIFAR-100 Tiny-ImageNet DomainNet
Benchmarks: CIFAR-100 Tiny-ImageNet
Metrics: test accuracy (last task) average accuracy over tasks
Baselines: EWC LwF iCaRL L2P TARGET GLFC LGA PILoRA InfLoRA FedAvg FedStack
Builds on: LoRA InfLoRA FedStack L2P DINO
Closest papers by meaning
- pFedClub: Controllable Heterogeneous Model Aggregation for Personalized Federated Learning (NeurIPS 2024 · similarity 0.89)
- FedAWA: Adaptive Optimization of Aggregation Weights in Federated Learning Using Client Vectors (CVPR 2025 · similarity 0.88)
- Fed-QSSL: A Framework for Personalized Federated Learning under Bitwidth and Data Heterogeneity (AAAI 2024 · similarity 0.88)
- PFedCS: A Personalized Federated Learning Method for Enhancing Collaboration among Similar Classifiers (AAAI 2025 · similarity 0.88)
- FedLMT: Tackling System Heterogeneity of Federated Learning via Low-Rank Model Training with Theoretical Guarantees (ICML 2024 · similarity 0.87)
- A Bayesian Approach for Personalized Federated Learning in Heterogeneous Settings (NeurIPS 2024 · similarity 0.87)
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.