Priority-Aware Learning-Unlearning Correction for Dynamic Decentralized LoRA Fine-Tuning

The article presents a priority-aware learning-unlearning correction framework for decentralized federated learning (DFL) using an orthogonal LoRA mechanism, addressing the challenges of dynamic edge networks where devices frequently join or leave. This framework allows for history-free updates by providing post-training contribution coordinates, enhancing the system's ability to adaptively correct fine-tuned parameters. The proposed system includes a resource allocation algorithm to optimize communication under constraints, demonstrating effective post-event corrections through experimental validation, which is crucial for practitioners aiming to implement efficient and adaptive DFL in real-world applications.