Research map / Federated learning and privacy / Personalized and heterogeneous FL
ConFREE: Conflict-free Client Update Aggregation for Personalized Federated Learning
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 ConFREE, a server-side conflict-free aggregation that projects conflicting client updates and searches a ball around the guidance vector to maximize the worst client's improvement, plugged into eight pFL methods.
Problem
Negative transfer in personalized federated learning, which the paper attributes to conflicting client updates (negative dot products) during server aggregation that bias the global model toward clients with more data.
Approach
Build a conflict-free guidance vector by projecting away pairwise-conflicting components of client updates, then pick an aggregation update within a ball around it that maximizes the worst client's loss reduction (solved via a dual over simplex weights), as a model-agnostic server-side complement to client-side pFL methods.
Key results
- ConFREE improves all eight pFL baselines in Table 1, with gains up to 1.92% (APFL on Flowers102 pathological setting).
- With 60 clients and 30% participation on Flowers102, APFL gains reach 3.42% (CIFAR100) and 3.42%-level improvements, and FedALA gains up to 2.79% on CIFAR100.
- Reported computation overhead is nearly unchanged versus original aggregation methods, with the paper stating acceptable overhead.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ConFREE aggregation to FedPAC | FedPAC · Flowers102, pathological heterogeneous | accuracy: 74.48 -> 74.68 | Helped |
| Add ConFREE aggregation to FedPAC | FedPAC · CIFAR100, pathological heterogeneous | accuracy: 72.60 -> 73.01 | Helped |
| Add ConFREE aggregation to FedPAC | FedPAC · CIFAR10, pathological heterogeneous | accuracy: 90.86 -> 91.42 | Helped |
| Add ConFREE aggregation to FedPAC | FedPAC · CIFAR100, practical heterogeneous (β=0.1) | accuracy: 61.94 -> 62.85 | Helped |
| Add ConFREE aggregation to FedPAC | FedPAC · TINY, practical heterogeneous (β=0.1) | accuracy: 46.63 -> 46.83 | Helped |
| Add ConFREE aggregation to FedPAC | FedPAC · Flowers102, practical heterogeneous (β=0.1) | accuracy: 68.56 -> 69.10 | Helped |
| Set c to 0.2 (ablation) | FedPAC + ConFREE · Flowers102, C=60, P=30% | accuracy: 58.95 -> 60.20 | Helped |
| Set c to 0.3 (ablation) | FedPAC + ConFREE · Flowers102, C=60, P=60% | accuracy: 59.04 -> 60.78 | Helped |
| Set c to 0.6 (ablation) | FedALA + ConFREE · Flowers102, C=60, P=30% | accuracy: 56.54 -> 59.33 | Helped |
| Set c to 0.8 (ablation) | FedALA + ConFREE · Flowers102, C=60, P=60% | accuracy: 58.71 -> 60.88 | Helped |
Surprising findings
- Gains are larger for methods that aggregate more global knowledge (APFL, FedALA, FedPAC) than for methods that keep more local layers (FedROD, FedRep, FedCP).
Limitations
- Projection guidance vector is only one choice; the authors note it could be replaced by other objectives.
- Hyperparameter c must be tuned per method and setting (optimal values vary from 0.1 to 0.8).
Experimental setup
Datasets: CIFAR10 CIFAR100 Tiny-ImageNet Flowers102
Benchmarks: CIFAR10 CIFAR100 Tiny-ImageNet Flowers102
Metrics: average test accuracy
Baselines: Per-FedAvg LG-FedAvg APFL FedROD FedRep FedCP FedALA FedPAC
Builds on: FedAvg FedALA FedCP FedPAC PCGrad
Closest papers by meaning
- FedCALM: Conflict-aware Layer-wise Mitigation for Selective Aggregation in Deeper Personalized Federated Learning (CVPR 2025 · similarity 0.90)
- FedAS: Bridging Inconsistency in Personalized Federated Learning (CVPR 2024 · similarity 0.88)
- Layer-wise Update Aggregation with Recycling for Communication-Efficient Federated Learning (NeurIPS 2025 · similarity 0.88)
- Efficient Federated Learning against Heterogeneous and Non-stationary Client Unavailability (NeurIPS 2024 · similarity 0.87)
- Fair Federated Learning under Domain Skew with Local Consistency and Domain Diversity (CVPR 2024 · similarity 0.87)
- Communication-Efficient Federated Learning with Accelerated Client Gradient (CVPR 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.