Surveillance video keyframe retrieval and attribute search have many application scenarios in traffic, security, education and other fields. The application of deep learning model to process massive video data to a certain extent alleviates manpower consumption, but it is characterized by privacy disclosure, large consumption of computing resources and long time. Based on the above scenarios, this study proposes a safe and fast video retrieval model for mass surveillance video. In particular, according to the characteristics of large computing power in the cloud and small scale of computing power in the surveillance camera, heavyweight model is deployed in the cloud, and the proposed tolerance training strategy is used for customized knowledge distillation, the distilled lightweight model is then deployed inside a surveillance camera, at the same time using local encryption algorithm to encrypt sensitive to image part, combined with cloud TEE technology and user authorization mechanism, privacy protection can be achieved with very low resource consumption. By reasonably controlling the "tolerance" of distillation strategy, the time-consuming of camera video input stage and cloud retrieval stage can be balanced, and extremely low retrieval delay is ensured on the premise of extremely high accuracy. Compared with traditional retrieval methods, the proposed model has the characteristics of security, efficiency, scalability and low latency. Experimental results show that the proposed model provides 9×-133× acceleration compared with traditional retrieval methods on multiple open data sets.