Communication-Efficient Federated Transfer Learning with Differential Privacy

Abstract

Federated learning (FL) enables distributed clients to train a shared model without centralizing raw data, but the exchanged gradients or model updates can still leak private information through gradient inversion attacks. Differential privacy (DP) mitigates this risk by clipping client updates and injecting calibrated noise before transmission. However, DP-protected FL still relies on repeated global communication rounds, which can be costly for resource-constrained deployments. To reduce this burden, federated transfer learning can reuse a source-trained initialization and adapt only selected trainable layers, lowering adaptation and communication costs. This paper proposes a communication-efficient federated transfer learning framework with client-side DP, where source clients first provide a transferable initialization and target clients then adapt the model with early convolutional layers frozen, while DP clipping and noise protect the uploaded updates during the DP-enabled adaptation stage. This design combines communication reduction through layer freezing with update privacy protection during target adaptation. In the verified 10-client setting, the 3-source-client FT+DP configuration reduces total system communication by 38.21% relative to baseline FL. Evaluation on the EuroSAT RGB classification task further shows a setting-specific trade-off among accuracy, communication cost, and privacy protection.