Federated Cloud Intelligence: A Privacy-Preserving, Trustworthy and Sustainable Framework for Multi-Cloud AI

Abstract

This study proposes a Federated Cloud Intelligence for Privacy-Preserving AI, with new layered framework that can support secure and eco friendly learning across different cloud providers. Instead of centralizing data, our method trains models locally on varied client datasets and combines their updates using federated learning (FL) to stay compliant with data protection rules. The experiment have shown that the federated setup reached an average accuracy of 0.844 over five communication rounds, just slightly lower than the centralized baseline of 0.850. Meanwhile, the loss decreased from 0.367 to 0.285, coming close to the centralized value of 0.318. To build trust, a blockchain-based layer that permanently stored updates with little extra cost, adding blocks each round with an average consensus delay of 0.189 seconds. Tests showed that this consensus process reduced the impact of malicious client attacks, keeping accuracy stable around 0.827. Further it is then incorporated with zero-knowledge proofs (ZKM), where adds only 0.196 seconds of latency and 260–360 MB GPU memory overhead and showcases an accuracy up to 0.844. A reinforcement learning agent optimized workload scheduling by shifting the computation from AWS to GCP, reducing carbon scores by 20% with minimal accuracy trade-off. Finally, explainability analysis revealed balanced provider contributions from 0.021 to 0.023 and highlighted key features such as logPurchases and storePurchases.