Our PhD Student Amr Abourayya under the supervision of Michael  Kamp (IKIM) successfully published his paper entitled „Little is Enough: Boosting Privacy by Sharing Only Hard Labels in Federated Semi-Supervised Learning“ at the top-tier AAAI 2025 Conference taking place in Pennsylvania (USA) from February 25 to March 4, 2025!

This work introduces a novel Federated Co-Training (FEDCT) approach that enhances privacy in federated learning by sharing only hard labels on public unlabeled datasets, enabling broader applicability and improved model quality across diverse tasks, including fine-tuning large language models.

Key Highlights:
– Enhanced Privacy: Empirically demonstrated protection against membership inference attacks.
– Flexibility: Supports interpretable models like decision trees and XGBoost.
– Efficiency: Reduces communication overhead.

Check out the Preprint here!