SARO: A Secure Adaptive Resource Orchestrator for Intelligent Cloud Management Using Machine Learning and Reinforcement Learning

Cloud platforms face challenges in dynamically managing resources while ensuring security for hosted applications. Reactive resource allocation often leads to latency spikes or resource wastage, and cyber-attacks or anomalies can further degrade performance. This paper presents Secure Adaptive Resource Orchestrator (SARO), a framework that integrates machine learning (ML) prediction, anomaly detection, and reinforcement learning (RL) for proactive cloud resource orchestration. SARO's five-layer architecture combines an LSTM-based workload predictor, an autoencoderbased anomaly detector, and a Q-learning orchestrator, and can be extended to advanced RL methods such as Deep Q-Network (DQN), Proximal Policy Optimization (PPO), and SARSA for complex decision spaces. We implemented SARO using opensource tools including OpenStack, Kubernetes, TensorFlow, and PyTorch, and evaluated it using Google cluster workloads and the UNSW-NB15 security dataset. Enhanced with statistical analysis and semi-intelligent Kubernetes Horizontal Pod Autoscaler (HPA) baseline, results from ten independent trials show that SARO achieves$78 \pm 4.2 \%$CPU utilization and$130 \pm 8$ms latency, outperforming the Dynamic baseline ($70 \pm 3.9 \%$,$150 \pm 20$ms) and HPA$(68 \pm 4.4 \%, 155 \pm 25 ms)(\mathrm{p}< 0.05)$. SARO maintains 95% anomaly-detection accuracy with 5% false positives and zero SLA violations, confirming statistically significant improvements in efficiency, responsiveness, and security compared to conventional approaches.