Communication-Efficient Federated Domain Generalization in Edge Networks via Dynamic Prototype Regulation

Key points

• Federated learning (FL) is a distributed paradigm for Internet of Things (IoT) edge networks.1
• However, deploying Vision-Language Models (VLMs) over wireless edge networks presents distinct challenges: large model parameters increase uplink transmission latency over bandwidth-constrained fading channels, and non-independent and identically distributed (Non-IID) data exacerbates client drift.1
• To mitigate client drift, we introduce the Gradient-Conflict Aware Modulation (GCAM) mechanism.1
• Experiments demonstrate that DPGR-GCAM improves cross-domain generalization and sustains convergence under network fading scenarios with a 10% client participation rate.1

Sources (1)

• [1]Communication-Efficient Federated Domain Generalization in Edge Networks via Dynamic Prototype Regulation
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 05:19 AM

  To address these issues, we propose DPGR-GCAM, a communication-efficient federated domain generalization framework.

  Federated learning (FL) is a distributed paradigm for Internet of Things (IoT) edge networks.

Extractive summary: sentences quoted from the sources.