Robust Personalized Federated Continual Learning via Explainable Multi-Granularity Prompt.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Science

Background:

• Personalized Federated Continual Learning (PFCL) faces challenges in integrating temporal and spatial knowledge while ensuring client-specific performance.
• Existing Personalized Federated Learning (PFL) and Federated Continual Learning (FCL) methods overlook multi-granularity knowledge representation, leading to Spatial-Temporal Catastrophic Forgetting (STCF).
• Current approaches often rely on local client-side personalization, increasing computational load and vulnerability to malicious clients.

Purpose of the Study:

• To propose FedMGP+, a novel framework addressing the limitations of existing PFCL methods.
• To leverage multi-granularity prompts for overcoming STCF and enabling coarse-to-fine personalization.
• To reduce computational overhead and enhance security against malicious clients.

Main Methods:

• Utilizing coarse-grained global prompts for efficient shared knowledge transfer without spatial forgetting.
• Employing fine-grained local prompts for personalized knowledge learning to overcome temporal forgetting.
• Implementing Personalized Selective Prompt Fusion on the server to generate client-specific prompts, reducing local computation and mitigating attacks.

Main Results:

• Coarse-grained prompts guide regional attention, while fine-grained prompts enrich object-level representations.
• FedMGP+ effectively mitigates forgetting and significantly enhances personalized performance across clients.
• The framework demonstrates robust defense against poisoning and gradient leakage attacks.

Conclusions:

• FedMGP+ offers an effective solution for PFCL by utilizing multi-granularity prompts.
• The proposed method balances knowledge integration, personalization, and security.
• FedMGP+ represents a significant advancement in secure and efficient personalized federated learning.