A Comparative Study of Deep Neural Networks and Support Vector Machines for Unsupervised Anomaly Detection in Cloud Computing Environments
Ahmed Hassan
Department of Computer Science and Engineering, Assiut University, Egypt
Tamilselvan Arjunan
Keywords: Anomaly detection, unsupervised learning, deep learning, support vector machines, cloud computing
Abstract
Cloud computing has become ubiquitous, providing convenient, on-demand access to computing resources. However, the complexity of cloud environments makes them prone to faults and anomalies that can severely impact service quality. Unsupervised anomaly detection, which does not require labelled data, is thus essential for cloud providers to identify issues proactively. This paper presents a comparative study of deep neural networks (DNNs) and support vector machines (SVMs) for unsupervised anomaly detection in cloud environments. We evaluate the performance of autoencoders, LSTM-based models, one-class SVMs, and isolation forests on benchmark datasets from cloud providers. Our results indicate that while shallow autoencoders are insufficiently expressive, LSTMs and convolutional autoencoders with dimensionality reduction can capture cloud workload patterns effectively. SVMs match or outperform autoencoders, with one-class SVMs showing robust performance across workloads. Isolation forests underperform on seasonal cloud data. Overall, one-class SVMs provide the best option for accurate, low-latency anomaly detection. Our findings provide guidance to cloud providers on selecting suitable unsupervised learning models based on their performance, interpretability, and computational overhead. The comparative methodology and results will inform future research on adapting unsupervised learning for cloud anomaly detection.
Author Biography
Tamilselvan Arjunan
