A new paper proposes Personalized Federated Direct Preference Optimization (pFedDPO), letting each client keep a small local LoRA adapter on a shared base model to better fit heterogeneous…

#FederatedLearning #LoRA #PreferenceOptimization #PrivacyPreservingML
https://arxiv.org/abs/2610.12885

Fed-GRPO proposes a federated extension of Group Relative Policy Optimization for LLM fine-tuning, using GRPO's own reward statistics as zero-cost signals to drive aggregation, local training, and communication.…

#FederatedLearning #LLM #RLHF #PrivacyPreservingML
https://arxiv.org/abs/2610.11502

New arXiv work introduces Cocoon, a system architecture that improves differentially private training by managing correlated noises, addressing accuracy loss and overhead in large models…

#DifferentialPrivacy #MachineLearning #AIInfrastructure #PrivacyPreservingML
https://arxiv.org/abs/2510.07304

The first paper in this session was Scheu-Hachtel & Zalonis's "Enhancing Noisy Functional Encryption for Privacy-Preserving Machine Learning" on flexible, efficient private ML. (www.acsac.org/2025/p...) 2/6
#FunctionalEncryption #PrivacyPreservingML

UTrace, an audit framework for privacy‑preserving ML, was tested on four datasets (vision, NLP, malware) across ten poisoning attacks, spotting malicious participants in most runs. Read more: https://getnews.me/utrace-user-level-poisoning-for-private-machine-learning/ #privacypreservingml #utrace