Abstract: To improve the cracking difficulty of encrypted images and not significantly increase the decryption time, a method to protect image data was proposed in this paper. First, a dynamic cipher verification method was proposed, which expanded the number of cipher text and ensured that the cipher text was different each time when typing the same plaintext so as to prevent the cipher algorithm from being cracked by dictionary or exhaustive method. At the same time, the encryption and decryption performance was adjusted independently according to the computer system environment. Second, the Rubik's cube algorithm was proposed, which the pixel and password data was rearranged into a hexahedral structure, the data on the face and bit was confused according to the form of cross axis, and the purpose of encrypting the image was achieved. According to the Rubik's cube principle, the data on each face was arranged in turn according to the password data sequence and pixel correlation. Experimental results show that this method can effectively prevent image privacy leakage and algorithm cracking, avoid the neural network to replay the pixel information, and run efficiently in the online image recognition system.