Open Access

ARTICLE

Adversarial-Resistant Cloud Security Using Deep Learning-Enhanced Ensemble Hidden Markov Models

Xuezhi Wen^1,2, Eric Danso^1,2,*, Solomon Danso¹

1 School of Computer Science, Nanjing University of Information Science and Technology, Nanjing, 210044, China
2 School of Cyber Science and Engineering, Nanjing University of Information Science and Technology, Nanjing, 210044, China

* Corresponding Author: Eric Danso. Email:

Journal of Cyber Security 2025, 7, 439-462. https://doi.org/10.32604/jcs.2025.070587

Received 19 July 2025; Accepted 19 September 2025; Issue published 17 October 2025

Abstract

Cloud-based intrusion detection systems increasingly face sophisticated adversarial attacks such as evasion and poisoning that exploit vulnerabilities in traditional machine learning (ML) models. While deep learning (DL) offers superior detection accuracy for high-dimensional cloud logs, it remains vulnerable to adversarial perturbations and lacks interpretability. Conversely, Hidden Markov Models (HMMs) provide probabilistic reasoning but struggle with raw, sequential cloud data. To bridge this gap, we propose a Deep Learning-Enhanced Ensemble Hidden Markov Model (DL-HMM) framework that synergizes the strengths of Long Short-Term Memory (LSTM) networks and HMMs while incorporating adversarial training and ensemble learning. Our architecture employs LSTMs for automated feature extraction from temporal cloud logs (such as Application Programming Interface (API) traces and network flows) and HMMs for interpretable attack state modeling, with an ensemble voting mechanism to enhance robustness. The framework is hardened against adversarial attacks through Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD)-based adversarial training, ensuring resilience against evasion attempts. Comprehensive experiments on the Canadian Institute for Cybersecurity Intrusion Detection System 2018 (CIC-IDS2018) and NSL-KDD datasets demonstrate that our approach achieves 92.4% accuracy (compared to 88.2% for a standalone LSTM) and maintains an F1-score of 0.82 under strong adversarial perturbations (ε = 0.2), outperforming state-of-the-art baselines, Convolutional Neural Network (CNN), Support Vector Machine (SVM), and HMM by 9.7%–15.8% in F1-score. The ensemble reduces the Adversarial Success Rate (ASR) compared to single models, while adding only 12% inference latency overhead. Statistical significance testing (p < 0.001) confirms these improvements. Key innovations include: a hybrid LSTM-HMM architecture for joint feature learning and state transition modeling, adversarially augmented training data to improve robustness, and majority voting across an ensemble of DL-HMMs to mitigate bias. This work advances cloud security by delivering a detection system that is accurate, interpretable, and adversarial-resistant qualities critical for real-world deployment.

---

Keywords

---