In current research, scholars have focused their efforts on employing machine learning methodologies to analyze multi-media data in order to classify searches made by users. In order to develop a dependable categorization algorithm that functions effectively inside a high-dimensional space, a combination of a robust classification and an extensive feature extractor is employed. In this paper, the integration of three distinct methods is employed to develop a probabilistic belief space policy. The techniques under consideration are assumption space planning, maximum entropy reinforcement learning (ML-RL), and generative adversary modeling (GAM). The offered techniques provide an assessment of different unmodeled adversarial techniques in order to attain robustness. This is motivated by the fact that the simulation include malicious behaviors. The aforementioned framework is employed with the intention of diminishing the agent's capacity for action dependability. The utilization of the reinforcement-based Deep Learning (DL) technique has the potential to be applied in the context of multi-model classification.