Convolutional neural network and wavelet composite against geometric attacks a watermarking approach

Digital e-governance has grown tremendously due to the massive information technology revolution. Banking, Healthcare, and Insurance are some sectors that rely on ownership identification during various stages of service provision. Watermarking has been employed as a primary factor in authenticating stakeholders in such circumstances. In this work, a three-layer feature-dependent image watermarking approach in the transform domain has been proposed. In this Discrete Wavelet Transform (DWT) influenced approach, the first decomposition level holds the Singular Values of a specific encrypted logo. In the second level of decomposition, the specific textual authentication signature is included in an arithmetic coding tag. The third level of decomposition has been utilised to keep the concerned identity of the owner in a compressed form using run-length coding. The proposed uniqueness of the scheme involves embedding a heavy payload watermark in the chosen grayscale cover image by utilising feature extraction and data compression, with a focus on preserving perceptual transparency and robustness. Various geometric variations and noise patterns, including Gaussian, salt and pepper, rotation, and cropping, were applied to the watermarked image to ensure its attack-resistant capability. After extracting the watermarks one by one, the reversibility of the cover image has been recovered through the Convolutional Neural Network (CNN) with a very low Mean Square Error (MSE). The proposed DWT-based scheme has achieved high perceptual transparency, as demonstrated by the Structural Similarity Index Measure (SSIM) and Normalised Correlation (NC) approach, which approaches unity. The integration of CNN enhances its robustness by recovering images against various attacks, as evidenced by achieving a PSNR of about 44 dB. Additionally, high SSIM (> 0.99) and NC (~ 1.0) values indicate enhanced perceptual transparency and robustness. The Bit error rate remained minimal post-attack, confirming reliable watermark recovery with CNN-aided restoration.