This paper introduces a new surveillance platform which is equipped with multiple parallel deep learning frameworks. The deep learning frameworks are used for the face recognition of input image and video streams from CCTV cameras in security applications. Each deep learning framework has its own accuracy (related to recognition performance) and operation time (related to system stability) those are in tradeoff relationship. Based on this system architecture, a new dynamic control algorithm which selects one deep learning framework for time-average security-level (i.e., machine learning accuracy for recognition and classification) maximization under the consideration of system stability. The performance of the proposed algorithm was evaluated and also verified that it achieves desired performance.