SFLaaS: Carbon-Constrained NAS Framework for Sustainable Federated Learning

A novel framework named Sustainable Federated Learning as a Service (SFLaaS) has been introduced to tackle sustainability issues within federated learning settings. This framework, outlined in a paper available on arXiv (2608.14359), features a carbon-constrained Neural Architecture Search (NAS) method aimed at accommodating diverse sustainability requirements from users. SFLaaS converts consumer sustainability profiles into a viable architecture space prior to federated execution, ensuring that only carbon-compliant architectures are evaluated. It also incorporates a mechanism for estimating carbon feasibility at the consumer level to assess potential architectures amid fluctuating carbon conditions. Furthermore, a sustainable scheduling strategy for consumers is employed to dynamically choose suitable participants and distribute local tasks, thereby enhancing consumer involvement and data coverage. This research is crucial for advancing environmentally friendly AI systems, especially in federated learning contexts where carbon emissions are increasingly significant. The paper can be found on arXiv under the identifier 2608.14359.