A known plaintext attack (KPA) algorithm based on ResNet is proposed for encryption system based on double random phase amplitude encoding (DRPAE). Our attack mainly uses a series of plaintext-ciphertext amplitude image pairs to make training set in residual network. Inputting plaintext ciphertext into neural network for training can make neural network fit the encryption process from ciphertext to plaintext. Independent of the ciphertext outside the training set， the trained network can effectively recover the original image， so as to attack the encryption system and decrypt the image. In this paper， the validity and feasibility of the attack method are analyzed theoretically and verified by computer simulation.