An Efficient Decentralized Federated Learning Framework Based on Directed Acyclic Graph

Federated Learning (FL) has gained significant attention and application in the field of Internet of Things (IoT). To address inherent vulnerabilities and trust risks in traditional centralized FL architectures, blockchain technology has been utilized to construct decentralized and trusted FL systems in the IoT domain. Nevertheless, conventional blockchain consensus mechanisms, such as Proof-of-Work (PoW), often impose substantial additional resource consumption. Furthermore, FL systems often involve large-scale model training, leading to considerable communication costs and storage pressures on the blockchain. This paper presents a novel lightweight, trusted, and decentralized FL architecture called DTFed, which is based on directed acyclic graph (DAG) ledger technology. DTFed exploits DAG ledger technology to establish an asynchronous FL system, leveraging the lightweight and efficient consensus mechanism of DAG to ensure system security. To mitigate communication costs while preserving effective FL performance, we adopt the concept of knowledge distillation. Soft labels are employed as intermediate results of Federated Learning instead of traditional model parameters, resulting in reduced communication costs. The experimental results demonstrate that DTFed enhances training efficiency, communication efficiency, and security compared to existing popular frameworks.