TinyFed6G: Federated Learning With TinyML for Resource-Constrained Intelligence in 6G Edge Networks

The rising prevalence of ultra-low-power microcontrollers in intelligent edge devices has created a demand for collaborative learning on devices. Federated Learning (FL) platforms are designed for edge devices; however, their functionality may be limited due to edge device constraints on memory, energy, and communication bandwidth, particularly in non-IID (Non-Independent and Identically Distributed) data distributions and heterogeneous hardware. To overcome these issues, this paper proposes TinyFed6G, a communication-efficient hierarchical FL framework that is 6G-aligned with TinyML (Tiny Machine Learning) clients. The architecture supports a dual-mode execution that enables the device to dynamically switch between training mode and inference mode based on real-time profiling of the device. Enhanced communication efficiency is achieved through semantic-aware compression, which filters quantized model updates based on a cosine similarity threshold. Personalized learning is accomplished by selectively fine-tuning the head layer of the model. The performance of the framework is evaluated using distributions. Experimental results demonstrate that TinyFed6G achieves a final accuracy of 86.2%, a 6.5% personalization gain.