Performance Evaluation of Autoencoder and JPEGBased Image Compression Schemes Integrated with AES and RSA Encryption

With the rapid growth of digital data production and transmission, the need to protect sensitive content and avoid unauthorized access has emerged. Data security is achieved through encryption, but the increasing volume of data requires efficient compression techniques to reduce bandwidth and storage expenses. In this paper, the complete framework for the evaluation of deep learning-based Autoencoder (AE) and traditional JPEG compression with symmetric (AES-128) and asymmetric (RSA-2048) cryptographic standards are proposed. A multi-faceted approach is utilized to analyze the performance of four different configurations: AE+AES, AE+RSA, JPEG+AES and JPEG+RSA. This approach includes visual quality (PSNR, SSIM), cryptographic security (NPCR, UACI, Entropy) and statistical randomness and hardware resource utilization. The experimental results on 52,000 high-resolution images from the FFHQ dataset show a large tradeoff between reconstruction fidelity and computational efficiency. JPEG based configurations have better visual reconstruction (PSNR ≈ 44.64 dB) and faster processing speed, but they are highly memory dependent (up to 3.9 GB RAM). In contrast, the proposed AE-based latent encryption method is very lightweight with a small memory footprint (0.01 GB RAM) which makes it an ideal solution for resource-constrained edge devices and IoT applications. The results showed that the deep learning-based compression combined with classical encryption provides a robust and flexible architecture for secure image transmission over different hardware platforms.