ML-Based Predictive Autoscaling on Aws

Traditional cloud auto-scaling systems rely on static threshold-based rules that react only after workload changes occur, leading to delayed responses, inefficient resource utilization, and increased latency. To address these limitations, this paper proposes a predictive auto-scaling system on AWS using machine learning and federated learning techniques. The system utilizes Long ShortTerm Memory (LSTM) models to analyze historical and real- time workload data, including CPU utilization, memory usage, and request rates collected through Amazon CloudWatch. Unlike centralized approaches, federated learning is employed to train models across distributed cloud nodes without sharing raw data, ensuring data privacy and reducing communication overhead. The locally trained models share updates that are aggregated to form a global predictive model capable of accurately forecasting future workload demands.Based on these predictions, the system proactively scales AWS EC2 instances using Auto Scaling Groups, enabling timely resource allocation before performance degradation occurs. This approach improves application responsiveness, reduces latency, optimizes resource utilization, and lowers operational costs. Overall, the proposed system provides a scalable, efficient, and privacy- preserving solution for intelligent cloud resource management.