Zero Trust Cloud Security Using Federated Graph Neural Networks (Fed-GNN)

The distributed and multi-tenant nature of many cloud computing environments means they are more vulnerable to advanced cyber threats, yet traditional, centralized intrusion detection systems will not satisfy the privacy conditions required with Zero-Trust security-based models. It's still tough to detect advanced attacks, like lateral movements and insider threats, if security analytics depend on data analysis confined in silo or centralized. In this paper, we propose a Zero-Trust cloud security system based on Federated Graph Neural Networks (Fed-GNN) to solve these limitations. The main objective of the proposed method is to enable privacy-preserving and scalable intrusion detection that collaboratively learns graph-based attack patterns among cloud tenants without requiring the sharing of raw data. Every tenant builds interaction graphs locally from network and system logs and trains a Graph Neural Network, with encrypted model updates being the only exchanged elements during federated aggregation. The model is combined with Zero-Trust policy engine to accomplish ongoing risk-based access control. Experimental results on real-world cloud intrusion datasets show that the proposed Fed-GNN can achieve accuracy of 97.1%, F1-score of 95.8% and reduce the false positive rate to 3.1 %, which is better than traditional machine learning and centralized deep learning baselines by$\mathbf{3 \%} \boldsymbol{-} \mathbf{5 \%}$. It also converges efficiently in terms of fewer communication rounds, demonstrating the effectiveness and practicability of Fed-GNN for secure practical cloud environment.