CONTEXT-AWARE AND ANOMALY-RESILIENT DEEP LEARNING FOR ADAPTIVE RESOURCE ORCHESTRATION IN DYNAMIC CLOUD ENVIRONMENTS

Abstract

Cloud environments change continuously as request rates, application dependencies, container placement, capacity, and failure conditions vary. Conventional resource controllers often react to lagging signals such as CPU utilization. This can lead to delayed scaling, unnecessary capacity, or service-level objective (SLO) violations. Predictive deep learning can estimate future demand, while anomaly detection can help distinguish ordinary workload changes from faults, attacks, and telemetry errors. However, these functions are often developed separately, leaving the orchestration policy without a consistent view of context and uncertainty. This paper proposes a context-aware and anomaly-resilient framework that combines temporal workload representation, context fusion, anomaly scoring, constrained policy learning, and a safety layer. Context includes workload history, service objectives, application dependencies, resource limits, queue length, recent deployments, and infrastructure health. The policy can select scaling, resource resizing, placement, traffic management, or conservative fallback actions. A multi-objective reward balances latency, SLO compliance, cost, energy use, and action stability. The paper specifies an evaluation protocol using trace-driven replay, controlled anomaly injection, ablation tests, and comparisons with reactive and forecast-only baselines. The accompanying numerical chart is synthetic and is provided only to demonstrate how results could be reported; it is not an empirical finding. The contribution is therefore an architecture and a testable research plan, not a claim of validated production performance. The paper also discusses data quality, distribution shift, explainability, safety, privacy, and deployment governance.