PrivAudit: A Federated Transformer-Driven Anomaly Detection Framework with Differential Privacy for Secure Cloud Infrastructures
Nellore Kapileswar, Judy Simon · 2025
The quick rise in cloud computing has resulted in many security problems, mostly concerning the safety of sensitive data from advanced cyber-attacks. Conventional ways of detecting anomalies are unable to preserve privacy and scale well enough for today's cloud environments. To solve these problems, this paper introduces PrivAudit, a new Anomaly Detection Framework that uses Federated Transformers and protects data with Differential Privacy. Using Transformers' strong ability to analyze sequences, PrivAudit successfully captures both time-based and location-based patterns in network traffic to spot unusual actions. With Federated Learning, data stays local on each client's device which greatly limits privacy concerns and allows the use of collective training for models. In addition, differential privacy is used to hide confidential data when aggregating gradients, giving formal protection against attempts to infer secret information. Many experiments on standard cloud security datasets show that PrivAudit is more effective at detecting faults, causing fewer mistakes and preserving privacy than existing approaches, so it is a good fit for the next generation of secure cloud networks.