A dual-layer privacy-preserving federated learning framework

Abstract

With the exponential growth of personal data use for machine learning models, significant privacy challenges arise. Anonymisation and federated learning can protect privacy-sensitive data at the cost of accuracy but there is lack of research on hybrid approaches. This paper uses federated learning and traditional centralised machine learning to evaluate the effectiveness of different anonymization strategies in environments with independent and identically distributed data. It considers the two layers of data collection (layer one) and model training (layer two) on three scenarios: (i) local data collection and local anonymisation for federated model training, (ii) central data collection before anonymisation for centralised model training, and (iii) central aggregation of locally anonymised data for centralised model training. Our assessment shows that the performance of the models generally decreases with increasing anonymity constraints, but the extent of decrease varies across different scenarios. In addition, we propose a dual-layer federated learning framework that applies differential privacy to ensure privacy during both data collection and model training stages. Evaluation on real-world datasets demonstrates that our framework achieves both acceptable data anonymization and model accuracy.

More information

Identifiers

ORCID for Thanassis Tiropanis: orcid.org/0000-0002-6195-2852

Catalogue record

Export record

Altmetrics