Dynamic pixel-wise hybrid multi-chaotic image encryption with DCT steganography and CNN-assisted restoration

Abstract Image encryption has become an essential requirement for protecting sensitive visual information, particularly in medical imaging, military communications, cloud storage, and multimedia transmission systems, where unauthorized access and cyberattacks continue to increase rapidly. However, many existing chaos-based image cryptosystems still suffer from several limitations, including weak randomness, insufficient resistance against statistical and differential attacks, fixed chaotic initial conditions, high computational complexity, and limited robustness against image-processing attacks such as filtering and histogram equalization. In addition, several conventional approaches focus solely on encryption without integrating secure steganographic protection mechanisms. To address these challenges, this paper proposes a novel hybrid multi-chaotic image cryptosystem combined with DCT-based steganography for secure image encryption and concealment. The proposed framework employs cosine-square chaotic modulation integrated with logistic, cubic, tent, Bernoulli, and quadratic chaotic maps to generate highly dynamic pixel-wise encryption keys with extreme sensitivity to initial conditions. Unlike traditional static chaotic systems, the proposed method generates distinct chaotic key streams for every image pixel according to the image dimensions, significantly improving randomness, diffusion, and unpredictability. Furthermore, the encrypted image is hidden inside a cover image using the Discrete Cosine Transform (DCT), providing an additional security layer against visual and statistical attacks. To improve the quality of the reconstructed images after decryption, adaptive enhancement and denoising stages using Gaussian filtering, median filtering, histogram equalization, and CNN-assisted restoration are incorporated. Extensive experimental and security analyses demonstrate the effectiveness and robustness of the proposed cryptosystem. The obtained results achieved high entropy values reaching 7.99, NPCR values up to 99.99%, UACI values exceeding 33%, and correlation coefficients approaching zero, confirming excellent statistical randomness and strong resistance against differential attacks. Moreover, the generated chaotic sequences successfully passed the NIST statistical randomness tests and Chi-square analysis, demonstrating superior cryptographic randomness characteristics. Additional security evaluations including chosen-plaintext attacks, known-plaintext attacks, filtering attacks, histogram equalization attacks, Gaussian noise attacks, and cropping attacks further verified the robustness and stability of the proposed framework. Comparative analyses with several recent state-of-the-art encryption techniques confirmed that the proposed method provides enhanced security performance, improved attack resistance, high key sensitivity, and efficient computational complexity, making it highly suitable for secure real-time image communication applications.