Federated & Private ML

Multi-Objective Privacy-Utility Optimisation in Federated Learning

Multi-Objective Optimization for Privacy-Utility Balance in Differentially Private Federated Learning

Instead of guessing a clipping norm for private federated learning, let an optimiser balance privacy and accuracy as training goes.

The problem

Clip too loosely and privacy suffers; clip too tightly and the model degrades. Picking one fixed norm is guesswork.

The approach

An adaptive clipping mechanism driven by a multi-objective optimisation that weighs privacy and utility together, with convergence analysis.

The outcome

Consistently beat fixed-clipping baselines on MNIST, Fashion-MNIST and CIFAR-10 under the same privacy budget.

In detail

Formulated differential privacy in federated learning as a multi-objective optimisation problem. Instead of using a fixed clipping norm, the method adapts clipping during training while explicitly considering both privacy and utility. Experiments across standard vision datasets show the value of adaptive clipping when the objective is to retain model performance under the same privacy constraints.

Built with

• Differential Privacy
• Optimisation
• Federated Learning

Source

• Multi-Objective Optimization for Privacy-Utility Balance in Differentially Private Federated Learning, arXiv:2503.21159, 2025

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: Meta-Clip: Privacy-Preserving Few-Shot Learning

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