Carbon Efficient Federated Learning Algorithms Via Adaptive Client Participation

Abstract

The proliferation of decentralized data has made Federated Learning (FL) the preferred paradigm for privacy-preserving distributed artificial intelligence. Nonetheless, the significant amount of computation and communication overhead involved in training advanced models in thousands of edge clients results in significant carbon emissions and environmental harm. In this paper, an algorithmic design is introduced to implement carbon-efficient federated learning via a process referred to as adaptive client participation. Contrary to conventional federated learning algorithms, which select edge clients based on network availability and the size of available data at the edge, the presented model uses adaptive client participation to change the probability of client selection based on real-time carbon intensity, local energy expenditure, and learning status. The experiments were performed on a distributed dataset with high diversity to emulate non-IID edge networks. From the simulations, one can clearly observe the environmental benefit obtained from the adaptive participation algorithm since it has been found that it significantly cuts down the carbon emissions associated with the operation of the system by 34.2% while maintaining high accuracy of the global model at 92.5%. Moreover, the algorithm has also proved to be highly efficient in terms of convergence since it requires 22.1% fewer iterations than the baseline carbon-unaware FL models to achieve the required level of accuracy. This is possible since only the use of edge resources is enabled during periods when the power grid uses renewable sources of energy. The paper thus provides a guideline for implementing green computing solutions in edge intelligence.