AI-Based Anomaly Detection in Multi-Cloud Traffic Using Deep Learning Models

Abstract

The rapid adoption of multi-cloud architectures across enterprises introduces significant challenges in ensuring consistent and reliable security monitoring.The dynamic and heterogeneous nature of traffic across different cloud platforms such as AWS, Azure, and Google Cloud makes anomaly detection complex and prone to high false-positive rates when using traditional rule-based or statistical methods. This paper presents an AI-driven anomaly detection framework that leverages deep learning models, particularly Long Short-Term Memory (LSTM) networks and autoencoders, to identify anomalous patterns in real time multi-cloud traffic. I design a simulation-based pipeline to generate labeled multi-cloud traffic, incorporating normal behaviors and diverse attack scenarios, including zero-day threats. Multiple deep learning architectures are trained and evaluated using precision, recall, F1-score, and inference latency. I proposed hybrid LSTM-Autoencoder model outperforms baseline methods, achieving an F1-score of 92.6% in detecting subtle and previously unseen anomalies, while maintaining low latency suitable for real-time inference.I present a case study of deploying the model in a realistic enterprise environment with federated cloud infrastructure. The results demonstrate the system’s effectiveness in detecting distributed denial-of-service (DDoS), data exfiltration,and lateral movement attacks with minimal overhead. This research advances the field of intelligent cloud security by introducing a scalable, adaptive, and efficient framework for anomaly detection in multi-cloud environments, supporting proactive threat mitigation in modern digital infrastructures.