DP-SimAgg: Privacy-Preserving Federated Learning for Brain Lesion Segmentation
A new federated learning framework called DP-SimAgg has been developed by researchers to improve both privacy and efficiency in medical imaging, particularly for segmenting brain lesions. This framework tackles two significant issues in federated learning: the varied data distributions among different institutions and the risk of information leakage during model updates. DP-SimAgg combines similarity-weighted aggregation with a differential privacy mechanism on the server side. It employs L2 clipping to limit updates from collaborators, calculates aggregation weights based on similarity to address non-IID data distributions, and adds calibrated Gaussian noise at the central server, ensuring privacy guarantees for each round under a specified sensitivity limit. The framework utilizes Intel's OpenFL platform and has been tested on the FeTS 2022 dataset. The findings are published in a paper on arXiv (arXiv:2608.00872).
Key facts
- DP-SimAgg is a privacy-preserving federated learning framework for brain lesion segmentation.
- It integrates similarity-weighted aggregation with server-side differential privacy.
- The method uses L2 clipping to bound collaborator updates.
- It computes similarity-based aggregation weights to mitigate non-IID data distributions.
- Calibrated Gaussian noise is injected at the central server for privacy guarantees.
- The framework is implemented using Intel's OpenFL platform.
- Evaluation was performed on the FeTS 2022 dataset.
- The paper is available on arXiv with ID 2608.00872.
Entities
Institutions
- Intel
- arXiv