FBID: Adaptive Personalized Federated Learning for Robust OOD Attack Detection in IoT
A recent study introduces Federated Bandit Intrusion Detection (FBID), an innovative personalized federated learning framework aimed at improving the detection of out-of-distribution (OOD) attacks within diverse IoT networks. This research, accessible on arXiv (2608.04073), highlights the shortcomings of current PFL techniques that depend on client-side self-optimization, which may result in excessive personalization and weakened OOD detection. FBID implements server-side personalization control through a contextual multi-armed bandit, adjusting each client's local training intensity based on performance and behavior. It also features a trust-based blending mechanism to generate client-specific interpolation coefficients between global and local models. This framework seeks to bolster resilience against OOD attacks in non-IID data contexts, making it crucial for IoT intrusion detection.
Key facts
- FBID is a novel adaptive PFL framework for intrusion detection in IoT.
- It addresses over-personalization in existing PFL methods.
- Uses a contextual multi-armed bandit at the server for personalization control.
- Introduces a trust-based blending mechanism for client-specific interpolation.
- Aims to improve OOD attack detection in non-IID data environments.
- Paper available on arXiv with ID 2608.04073.
- Announcement type is cross.
- Focuses on server-side personalization control.
Entities
Institutions
- arXiv