Federated learning method cuts communication for edge vision language models
Communication-Efficient Federated Domain Generalization in Edge Networks via Dynamic Prototype Regulation
Machine Learning
Summary
Sending large AI models from many small devices to a central server can be slow and unreliable, especially when internet connections vary and data are different across devices. The authors designed a new approach that smartly reduces the amount of data each device needs to send by adjusting what to share depending on connection quality. They also include a way to fix conflicts in learning updates from different devices, making the overall training more stable and faster even with limited participation. Their method helps AI models to better understand new types of data without needing all devices to communicate heavily.
What this means in practice
- •For iot network engineers: Design and deploy vision-language AI systems on bandwidth-limited edge devices using fewer transmissions and robust model updates under varying wireless conditions.
- •For mobile edge computing teams: Improve responsiveness and stability of federated AI services by dynamically adjusting data sharing based on channel quality and minimizing client dropouts.
Authors
Yuankun Xia, Hui Wang, Ling Yuan, Ziyu Hong, Di Yin, Yufeng Zhou
Abstract
Federated learning (FL) is a distributed paradigm for Internet of Things (IoT) edge networks. 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. Existing feature alignment solutions often introduce auxiliary computations that consume time slots allocated for wireless parameter transmission, increasing the likelihood of straggler dropouts under system delay constraints. To address these issues, we propose DPGR-GCAM, a communication-efficient federated domain generalization framework. By employing a dual-parameter-efficient fine-tuning (PEFT) architecture, we design a channel-aware tiered transmission mechanism that adaptively truncates payloads to maximize node survival rates under fading conditions. To mitigate client drift, we introduce the Gradient-Conflict Aware Modulation (GCAM) mechanism. GCAM operates within the low-dimensional feature manifold to dynamically resolve multi-objective optimization conflicts with minimal edge computation latency, thereby reducing computation-induced network outages. Experiments demonstrate that DPGR-GCAM improves cross-domain generalization and sustains convergence under network fading scenarios with a 10\% client participation rate.