Abstract: In order to solve the problems of chaotic system degradation in traditional single chaotic image encryption algorithm and the encryption algorithm is difficult to resist the chosen plaintext attack, a new 3D-Duffing chaotic system is constructed, and a plaintext related bit cyclic diffusion encryption algorithm is designed. The image is encrypted for the first time. RBF neural network is utilized to predict the Henon chaotic sequence and obtain the predicted key stream. A method of randomly embedding prediction key in the encryption sequence is proposed to encrypt the image twice, and it can resist the chosen plaintext attack effectively. The experimental results show that the algorithm can make the image pixel values completely change, and the encrypted image pixel values are evenly distributed, and the NPCR and UACI values are close to the theoretical values. The algorithm not only solves the above problems, but also further improves the security of the encryption algorithm, and has good application value.