Privacy-Driven Cloud AI: Federated Learning and Zero-Trust for Secure Multi-Domain Collaboration

This paper introduces the privacy-preserving cloud AI system, GWO-FLNet, which is an arrangement of Federated Learning (FL), Zero-Trust architecture (ZTA), and the Gray Wolf Optimizer (GWO) that allows secure and high-accuracy interaction between distributed and multi-domain systems. The first goal is to maximize the precision of the models and preserve the data privacy when the datasets of clients are statistically heterogeneous and domain-specific. GWO-FLNet utilizes a CNN-BiLSTM hybrid setup, which models spatial-temporal regularities and trains the message layer of a client without communication of raw signals between that client and the server. In order to address inter-client differences and training imbalance, GWO adjusts the learning rate, the size of BiLSTM units, and aggregation weights, personalized by the client during the federated training process. Access control and strict authentication are imposed through ZTA principles over all clients. The experimental data show that the advancement in the performance is significant as opposed to traditional FL models, including FedAvg and FedProx: average Dice Similarity Coefficient (DSC) increase of 7.8%, Intersection over Union (IoU) gain of 8.3%, and an increase in F1-score of 6.4%, besides 15% improvement in communication overhead. These results highlight the strength and scalability of the structure in situations with decentralized, nonuniform populations of data, which makes it highly appropriate for setups applicable in the real-time, privacy-conscious use of the cloud.