Federated Learning and Privacy in Large Language Models
Recent advancements in the field of Large Language Models (LLMs) have predominantly focused on enhancing efficiency, privacy, and robustness in federated learning environments. The general direction of the field is moving towards developing more sophisticated methods for parameter-efficient fine-tuning, model pruning, and privacy-preserving techniques, all while maintaining or improving model performance. Innovations in federated learning frameworks, such as the introduction of novel aggregation functions and adaptive mask expansion techniques, are paving the way for more efficient and secure model updates across distributed clients. Additionally, the integration of auxiliary models, like language models, into privacy-enhancing strategies is showing promising results in mitigating gradient inversion attacks and protecting sensitive data.
Noteworthy developments include:
- A federated learning framework for pruning LLMs that introduces a novel $\ell_0$-norm aggregation function.
- A method for approximating data ablations in LLMs through modular training and merging, significantly improving amortized training efficiency.
- A gradient attack algorithm tailored to spatiotemporal data that utilizes an auxiliary language model to enhance reconstruction accuracy.
- A subspace regularization method on LoRA structure that effectively balances model capacity and degree of forgetting.
- A novel approach to model merging in federated continual learning that ensures alignment of model responses across tasks and clients.