<![CDATA[The administration of assessment documents in The National Narcotics Board requires a mechanism that can conceal confidential information while preserving the visual quality of the carrier image. This study aims to develop an AI-assisted web-based steganography system for embedding assessment documents using a Binarized Neural Network (BNN)-based embedding region selection model. The proposed method extends conventional Least Significant Bit (LSB) steganography by incorporating BNN-based patch classification before data insertion. Assessment data submitted through a web form were automatically converted into PDF files and embedded into residency cover images using an adaptive LSB technique guided by the BNN model. Due to confidentiality restrictions, this study used nine valid residency images, which were segmented into 32×32-pixel patches, producing 900 patches categorized into suitable and unsuitable embedding regions. The BNN model achieved an accuracy of 91.7%, precision of 91.2%, recall of 92.2%, and F1-score of 91.7%. Image quality evaluation showed an average MSE of 2.722, PSNR of 43.79 dB, and SSIM of 0.981, indicating high visual similarity between cover and stego images. Functional testing using black-box testing was conducted on 12 system scenarios, including login validation, role-based access, PDF generation, image upload, embedding, extraction, usage history, and user management access. All 12 scenarios were successfully passed, resulting in a functional success rate of 100%. Implementation observation also indicated that direct conventional LSB embedding tended to produce larger stego-image file sizes than the proposed BNN-assisted adaptive LSB approach. The findings suggest that BNN-assisted region selection can support imperceptible document embedding in web-based steganography applications. However, the limited number of real images, the absence of formal steganalysis, and the lack of full quantitative baseline comparison remain limitations that should be addressed in future work.]]>
