Research map / Federated learning and privacy / Privacy attacks and secure learning
Breaking Data Silos in Parkinson’s Disease Diagnosis: An Adaptive Federated Learning Approach for Privacy-Preserving Facial Expression Analysis
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 an adaptive federated learning method that weights clients by gradient, data, and learning-efficiency contributions, and applies it to Parkinson's disease diagnosis from facial expressions using a newly collected PDFE dataset.
Problem
Parkinson's disease diagnosis from facial expressions is limited by small, siloed local datasets due to privacy and legal restrictions, and non-IID data across hospitals hurts federated aggregation.
Approach
Client aggregation weights are computed from three leave-one-out contribution signals: angular deviation between local and other-client gradients, the error of the others' aggregated model on a client's local data, and the relative loss-decrease rate; these are combined with fixed weights (0.5, 0.4, 0.1) and also local batch sizes are proportional to client data size.
Key results
- On PDFE PD diagnosis, the proposed method reaches 98.35% accuracy, versus 95.43% for ResNet18-DataAugment (the best baseline).
- On CIFAR-10 with Dirichlet a=0.1, the method achieves 72.87% vs 66.09% for FedSol (about 6.78 points higher).
- Under non-IID PD settings (a from 100 to 1 and image-lightness shifts), accuracy stays within 97.20-98.15%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use adaptive FL with gradient, data, and learning-efficiency contributions instead of FedSol | Simple-CNN · CIFAR-10 (a=0.1) | accuracy: 66.09 -> 72.87 | Helped |
| Use adaptive FL with gradient, data, and learning-efficiency contributions instead of FedSol | Simple-CNN · CIFAR-10 (a=0.5) | accuracy: 67.56 -> 71.45 | Helped |
| Use adaptive FL with gradient, data, and learning-efficiency contributions instead of FedAvg | Simple-CNN · CIFAR-10 (IID) | accuracy: 71.91 -> 72.95 | Helped |
| Use adaptive FL with gradient, data, and learning-efficiency contributions instead of FedSol | Simple-CNN · FMNIST (a=0.1) | accuracy: 89.83 -> 91.49 | Helped |
| Use adaptive FL with gradient, data, and learning-efficiency contributions instead of MOON | Simple-CNN · FMNIST (a=0.5) | accuracy: 89.91 -> 91.46 | Helped |
| Use adaptive FL with gradient, data, and learning-efficiency contributions instead of FedSol | Simple-CNN · FMNIST (IID) | accuracy: 92.14 -> 92.31 | Helped |
| Use adaptive federated learning instead of ResNet18-DataAugment | ResNet-18 · PDFE | accuracy: 95.43 -> 98.35 | Helped |
| Train across clients with varying data-scale distributions | Ours · PDFE (non-IID data scale, a=1, 10, and 100) | accuracy: 98.14 -> 97.78 | Mixed |
| Train with image lightness, contrast, and saturation heterogeneity | Ours · PDFE (image heterogeneity) | accuracy: 97.20; 97.31; 98.03 | Mixed |
| Add data and learning-efficiency contributions to gradient contribution (ablation) | Adaptive FL · PDFE | accuracy: 96.89 -> 98.35 | Helped |
| Add gradient and learning-efficiency contributions to data contribution (ablation) | Adaptive FL · PDFE | accuracy: 96.62 -> 98.35 | Helped |
| Add gradient and data contributions to learning-efficiency contribution (ablation) | Adaptive FL · PDFE | accuracy: 96.14 -> 98.35 | Helped |
Surprising findings
- Increasing local update rounds beyond 1 degraded performance, attributed to local overfitting.
- On CIFAR-10 at a=0.5, the proposed method (71.45%) is slightly below FedSol-free baselines' IID-like range, and FedSol scores lower than FedAvg at FMNIST a=0.5 (88.35 vs 89.88).
Limitations
- Dataset is small (95 PD patients) and collected from a single hospital, so multi-site generalization is not shown.
- Fixed contribution weights (lambda) were tuned by grid search on CIFAR-10 and reused for PD, with no sensitivity analysis on the PD task.
- The PD test set augments with elderly healthy data from Tsinghua-FED, which may introduce confounds such as age.
Experimental setup
Datasets: PDFE Oulu-CASIA RaFD CK+ Tsinghua-FED CIFAR-10 FMNIST
Benchmarks: PDFE CIFAR-10 FMNIST
Metrics: accuracy
Baselines: FedAvg FedProx MOON FedSol GLCM+SVM ResNet18-DataAugment DeiT-small EfficientNetV2 InceptionResnetV1
Builds on: FedAvg FedProx MOON FedSol Shapley value
Closest papers by meaning
- Accessible, At-Home Detection of Parkinson’s Disease via Multi-Task Video Analysis (AAAI 2025 · similarity 0.86)
- Care-PD: A Multi-Site Anonymized Clinical Dataset for Parkinson’s Disease Gait Assessment (NeurIPS 2025 · similarity 0.85)
- TULIP: Multi-camera 3D Precision Assessment of Parkinson's Disease (CVPR 2024 · similarity 0.82)
- OpticalDR: A Deep Optical Imaging Model for Privacy-Protective Depression Recognition (CVPR 2024 · similarity 0.81)
- MDD-5k: A New Diagnostic Conversation Dataset for Mental Disorders Synthesized via Neuro-Symbolic LLM Agents (AAAI 2025 · similarity 0.81)
- How Private is Private? A Comparative Study for Face De-Identification (NeurIPS 2026 · similarity 0.80)
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.