Enhancing Federated Learning Privacy Through Encrypted Computation
Abstract
Ensuring patient privacy while leveraging artificial intelligence in healthcare is a critical challenge. This work presents a privacy-preserving federated learning framework for classifying brain tumors from MRI scans. Without disclosing private patient information, the proposed system enables collaborative model training across multiple healthcare institutions. To prevent long-term key compromise and ensure secure transmission of intermediate CNN features, we employ homomorphic encryption using the CKKS scheme with periodic key rotation. Each institution locally trains a lightweight MobileNetV3-Small model and encrypts the extracted features with TenSEAL before transmitting them to a central server for aggregation. Our approach presents a novel encrypted feature sharing mechanism that improves privacy without sacrificing model performance, in contrast to traditional methods that encrypt gradients or model updates. Experimental evaluations on a brain MRI dataset demonstrate that the proposed framework increases the accuracy of models from 84.5% (traditional FL) and 87.1% (LEAF) to 94.1%. It safeguards sensitive healthcare data using full CKKS encryption with key rotation and achieves a 5× compression ratio, reducing communication overhead compared to LEAF’s 3×. These results highlight its potential for scalable, secure, and high-performance federated AI in real-world clinical settings
Read the paper
Related studies
- Active Learning during Federated Learning for Object Detection
- Fairness-aware federated learning
- Decentralized Federated Learning
- Building a Personally Identifiable Information Recognizer in a Privacy Preserved Manner Using Automated Annotation and Federated Learning
- Preservation of Patient Level Privacy: Federated Classification and Calibration Models
- Efficient explainable federated learning
Go deeper in Evidence Hub
Evidence Hub adds a study data sheet (design, sample size, evidence level, key numbers), an AI summary and a chat grounded in the paper's text.