Threat Detection in Cloud Computing Using Machine Learning

Abstract

Cloud computing has become a critical foundation for hosting large-scale applications and storing sensitive information, but its open and multi-tenant architecture exposes cloud infrastructures to a wide range of cyber threats. Traditional signature-driven intrusion detection systems provide limited defense against new and evolving attack strategies, creating a need for intelligent and adaptive threat detection mechanisms. This study presents a Transformer-based machine learning model enhanced with contrastive self-supervised training for cloud intrusion detection. Network-flow features are transformed into token sequences using embedding and positional encoding, enabling the model to capture relationships across heterogeneous traffic attributes. The proposed system is evaluated on large benchmark datasets, including CSE-CIC-IDS2018, ToN_IoT, and BoT-IoT. Experimental results show that the model achieves 99.21% accuracy, 98.87% macro F1-score, and 99.46% AUROC on CSE-CIC-IDS2018, outperforming existing machine learning baselines such as XGBoost, LightGBM, TabTransformer, and Bi-LSTM. Cross-dataset testing confirms strong generalization, with accuracy improving from 94.15% to 96.91% on ToN_IoT after limited fine-tuning. Low false-positive rates and millisecond-level inference latency indicate suitability for real-time deployment in cloud environments. The findings demonstrate that self-supervised Transformer architectures offer a scalable and effective solution for modern cloud threat detection.