Federated & Private ML

Meta-Clip: Privacy-Preserving Few-Shot Learning

Adaptive Clipping for Privacy-Preserving Few-Shot Learning: Enhancing Generalization with Limited Data

Few-shot learning on private data, without the usual drop in performance, by adapting how much each update is clipped.

The problem

Few-shot learning already has little data to work with; adding differential privacy usually costs even more accuracy.

The approach

Meta-Clip adapts clipping thresholds during training for DP-MAML, DP-Reptile and DP-MetaSGD, controlling disclosure while limiting overfitting.

The outcome

A better privacy-utility trade-off than existing privacy-preserving techniques across diverse benchmark datasets.

In detail

Introduced Meta-Clip, an adaptive clipping mechanism for privacy-preserving few-shot learning. The method is designed for differentially private meta-learning algorithms including MAML, Reptile and MetaSGD, dynamically adjusting the clipping threshold during training to better balance privacy, generalisation and learning capacity in small-data settings.

Built with

• Few-Shot Learning
• Differential Privacy
• Meta-Learning

Source

• Adaptive Clipping for Privacy-Preserving Few-Shot Learning: Enhancing Generalization with Limited Data, arXiv:2503.22749, 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: Few-Shot Medical Diagnosis with Model-Agnostic Meta-Learning

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