In this work, an improvement of the anomaly detection model for security surveillance cameras is introduced. Based on the state-of-the-art model, the knowledge distillation (KD) method is used for the transfer learning process. Three deep learning models were used and combined with the KD technique. The I3D was used for extracting embedding vectors from sequences of videos. The PEL4VAD and the UR-DMU models were used as the teacher and the student, respectively, in the KD method. The transfer learning process was performed on a publicly available dataset, namely the UCF-Crime, and our self-collected videos. The evaluation was implemented on the test dataset. It is shown that when utilizing the transfer learning process with the KD method, the accuracy of detection can be slightly improved to 91.76%. Since UR-DMU, the smaller network, is used for the inferring process, the computation is also more efficient.