Federated & Private ML

Utility-Enhanced Differentially Private Federated Learning

Federated Learning with Differential Privacy: An Utility-Enhanced Approach

Private federated learning that keeps more of its accuracy, by adding the privacy noise in a smarter place.

The problem

Federated learning alone does not guarantee privacy, and standard differential privacy adds enough noise to hurt model utility.

The approach

A Haar wavelet transform step plus a new noise-injection scheme that lowers the asymptotic bound on noise variance, with a full convergence analysis.

The outcome

Better model utility than existing approaches on real-world datasets under the same privacy guarantees.

In detail

Investigated the privacy-utility trade-off in differentially private federated learning and proposed a utility-enhanced training method based on Haar-wavelet transformation and a revised noise-injection strategy. The approach reduces the impact of privacy noise while retaining formal privacy protection, with convergence analysis used to characterise its behaviour relative to standard differentially private training.

Built with

• Federated Learning
• Differential Privacy
• Wavelets

Source

• Federated Learning with Differential Privacy: An Utility-Enhanced Approach, arXiv:2503.21154, 2025
• Improving the Utility of Differentially Private SGD by Employing Wavelet Transforms, 2023 IEEE International Conference on Big Data (BigData), pp. 1352–1359, 2023

More in Federated & Private ML

• IRT-FedLoRA: Reliability-Aware Federated Fine-Tuning
• ClusterGuardFL: Secure Cluster-Weighted Client Aggregation
• AR-FL: Anomaly-Resilient Federated Learning

Next project: Multi-Objective Privacy-Utility Optimisation in Federated Learning

By Kanishka Ranaweera, PhD, Machine Learning Engineer in Melbourne, Victoria, Australia. hello@kanishkabuilt.me