Research map / Federated learning and privacy / Personalized and heterogeneous FL
FedCALM: Conflict-aware Layer-wise Mitigation for Selective Aggregation in Deeper 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.
Identifies deep model degradation in personalized federated learning caused by server aggregation conflicts and proposes FedCALM, a layer-wise conflict projection plus QP-based conflict-aware aggregation that boosts SOTA PFL methods.
Problem
Personalized federated learning (PFL) performance degrades as base model depth increases because deeper layers are more sensitive to non-IID client data, causing server aggregation conflicts between client updates.
Approach
A layer-wise projection removes negative-cosine conflicts between client updates, and a QP-based conflict-aware mitigation step finds a global update close to the conflict-free average while guaranteeing a minimum positive contribution to every client; it works as a standalone aggregation method or as an add-on to existing PFL methods.
Key results
- On Flowers102 with ResNet4, FedCALM reaches 75.57% accuracy, 9.88% above the next-best method.
- Adding CALM to FedCP raises Flowers102 accuracy from 70.80% to 79.81% (+9.01%) under the practical non-IID setting.
- With N=60 clients on CIFAR100, FedCALM outperforms FedPAC by 6.75% and 4.10% at 30% and 60% participation respectively.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedCALM instead of the strongest baseline with ResNet4 | PFL · Flowers102 | accuracy: +9.88% | Helped |
| Use FedCALM instead of the strongest baseline with ResNet10 | PFL · Flowers102 | accuracy: +3.16% | Helped |
| Use FedCALM instead of the strongest baseline with ResNet18 | PFL · Flowers102 | accuracy: +1.59% | Helped |
| Use FedCALM instead of the strongest baseline with ResNet34 | PFL · Flowers102 | accuracy: +4.57% | Helped |
| Use FedCALM instead of the strongest baseline with ResNet4 | PFL · CIFAR100 | accuracy: +5.31% | Helped |
| Use FedCALM instead of the strongest baseline with ResNet10 | PFL · CIFAR100 | accuracy: +0.52% | Helped |
| Use FedCALM instead of the strongest baseline with ResNet18 | PFL · CIFAR100 | accuracy: +2.51% | Helped |
| Use FedCALM instead of the strongest baseline with ResNet34 | PFL · CIFAR100 | accuracy: +3.67% | Helped |
| Add CALM aggregation strategy to FedCP | FedCP with CNN · Flowers102 | accuracy: 70.80% to 79.81% (+9.01%) | Helped |
| Add CALM aggregation strategy to FedCP | FedCP with CNN · Flowers102 P | accuracy: 80.54% to 86.95% (+6.41%) | Helped |
| Add CALM aggregation strategy to FedCP | FedCP with ResNet10 · CIFAR10 | accuracy: 90.01 to 91.96 (↑1.95) | Helped |
| Add CALM aggregation strategy to FedALA | FedALA with ResNet10 · Flowers102 | accuracy: 63.70 to 72.41 (↑8.71) | Helped |
Surprising findings
- Accuracy of all four SOTA PFL methods declines as base model depth increases (deep model degradation), with FedPAC and FedALA falling below their CNN results.
- Adding CALM slightly reduces communication overhead while adding negligible computation overhead.
Limitations
- Standalone FedCALM underperforms FedCP+CALM on Flowers102 under the pathological setting.
- Evaluated only on image classification with CNN/ResNet base models and four datasets.
Experimental setup
Datasets: CIFAR10 CIFAR100 Flowers102 CINIC10
Benchmarks: CIFAR10 CIFAR100 Flowers102 CINIC10
Metrics: test accuracy conflict rate communication overhead computation time
Baselines: FedRep FedALA FedCP FedPAC
Builds on: FedRep FedALA FedCP FedPAC
Closest papers by meaning
- ConFREE: Conflict-free Client Update Aggregation for Personalized Federated Learning (AAAI 2025 · similarity 0.90)
- FedAS: Bridging Inconsistency in Personalized Federated Learning (CVPR 2024 · similarity 0.88)
- pFedClub: Controllable Heterogeneous Model Aggregation for Personalized Federated Learning (NeurIPS 2024 · similarity 0.88)
- PeFLL: Personalized Federated Learning by Learning to Learn (ICLR 2024 · similarity 0.88)
- Flexible Sharpness-Aware Personalized Federated Learning (AAAI 2025 · similarity 0.88)
- Fine-Tuning Personalization in Federated Learning to Mitigate Adversarial Clients (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.