<p>A novel, asymmetric image encryption-hiding scheme (AiEhS) using a reversible neural network (RNN) was developed, in which deep learning is employed to compress and hide a secret plain image (SPI), thereby enhancing the encryption efficiency and improving the hiding quality. First, AiEhS employs an auto-encoder to compress the SPI and designs a new encryption method for encrypting the compressed image to obtain a cipher image, reaching the first layer of encryption. Second, pixels in the cipher image are decomposed, combined, and scrambled to obtain another scrambled image. Thereafter, a trained RNN model is used to embed this scrambled image into a selected carrier image, resulting in a new carrier image hiding secrets, thus realizing the second layer of hiding. Moreover, AiEhS produces a pseudorandom sequence using a hyperchaotic map and constructs a new key model to achieve a plaintext dependency. The keys are then designed and distributed by the Rivest-Shamir-Adleman algorithm, effectively improving the security. Compared with traditional compressive-sensing-based image-hiding methods, the contributions of AiEhS are as follows: (1) A new scheme is designed using an auto-encoder to compress the SPI, which can reduce the time cost of both compression and reconstruction, thus accelerating the execution efficiency. (2) The scrambled image is hidden in a carrier image by RNN, which can increase the embedding amount and achieve better hiding quality. Furthermore, experiments show that AiEhS using deep learning can ensure better security and efficiency for image encryption and hiding, in contrast with the traditional image compression and embedding technique. For example, the peak signal-to-noise ratio for the reconstructed image exceeds 34 dB.</p>