Abstract: In order to balance the contradiction between camera using and privacy disclosure and address the problem of privacy disclosure from data sources,we proposed the visual privacy behavior recognition and protection algorithm for social robot based on image semantics,and developed a privacy protection system.Firstly,the scene recognition algorithm based on convolution neural network and attention model were proposed by introducing the attention model into the convolution neural network.Then the privacy-related images semantic description algorithm based on Inception-v3 and long short-term memory network was designed.Then the visual privacy behavior recognition and protection algorithm for social robot based on image semantics was formed.The indoor scene dataset MIT Scene-67 was used as the train and test datasets,and the comparison results with four algorithms show that the proposed algorithm is more superior in the aspect of recognition accuracy and robustness.The results show that the average privacy recognition accuracy and the privacy scene semantic description accuracy are 91.65%and 86.60%respectively,which indicate that the system has acceptable robustness and privacy protection performance.