Embedded devices, crucial components in various industries, often operate independently, executing specific tasks efficiently. Their compact size, low maintenance, and energy consumption make them highly desirable. With the ability to connect to networks, these devices facilitate communication with other devices, forming a robust computing system. However, image transmission on these devices poses a challenge, requiring a delicate balance between efficiency and cybersecurity. In this paper, we propose a novel solution, a depthwise separable convolutional autoencoder (DSCAE) network model, which is unique in its ability to address image compression and encryption simultaneously. This model incorporates the high-efficiency depthwise separable convolution (DSC) of the Xception network into the convolutional autoencoder (CAE) model, optimizing image transmission. It also utilizes the Xception middle flow structure to synthesize more features, thereby enabling the decoder to reconstruct the predicted image with greater accuracy and enhancing the model’s performance. The output of the encoder is in ciphertext format to ensure the confidentiality of transmitted images, effectively safeguarding them and reducing the risks associated with unauthorized access during image communication. The experimental results demonstrate the efficacy of this approach, with the original photos transmitted by the proposed deep learning image encoding method retaining image quality and encryption by transferring only one-sixtieth of the original image size. On the receiver site, the reconstructed images can achieve an average peak signal-to-noise ratio (PSNR) of 29 dB compared with the actual image at the transmitter, thereby significantly improving the efficiency and security of image transmission on embedded devices.