SacFL: Self-Adaptive Federated Continual Learning for Resource-Constrained End Devices.

Area of Science:

• Distributed computing and machine learning on edge devices.

Background:

• End devices generate dynamic data, challenging on-device machine learning models.
• Continual learning (CL) mitigates catastrophic forgetting but centralized CL raises privacy and data concerns.
• Federated CL (FCL) offers a privacy-preserving solution by keeping data local.

Purpose of the Study:

• To propose a novel FCL framework, self-adaptive federated CL (SacFL), addressing storage limitations, autonomous task shift detection, and adversarial task handling.
• To reduce storage demands for resource-constrained end devices.
• To enable autonomous triggering of CL or defense strategies.

Main Methods:

• An encoder-decoder architecture separates task-robust and task-sensitive components.
• Contrastive learning is used for autonomous data shift detection.
• The framework validates effectiveness on text and image datasets (Cifar100, THUCNews).

Main Results:

• SacFL significantly reduces storage requirements by retaining lightweight task-sensitive components.
• An autonomous mechanism effectively detects new tasks and distinguishes benign from adversarial ones.
• The framework demonstrates effectiveness in both class-incremental and domain-incremental scenarios.

Conclusions:

• SacFL provides a practical and efficient solution for continual learning on resource-constrained end devices.
• Autonomous capabilities enhance the adaptability and security of on-device models in dynamic environments.
• The developed demo system confirms the framework's real-world applicability.