Protecting machine-learning datasets with digital watermarking based on modified QR decomposition

Abstract

The paper examines an approach to protecting intellectual property rights for machine learning image datasets using digital watermarking based on QR decomposition. The proposed scheme involves embedding hidden identification information into the image structure through controlled modification of the elements of the upper triangular matrix obtained from QR decomposition, ensuring a combination of invisibility, robustness, and computational efficiency. To ensure numerical stability and efficiency in the watermark embedding and extraction processes, the method employs an improved QR decomposition algorithm with diagonal element correction, as well as a fast watermark extraction technique with linear computational complexity. The effectiveness of the approach is evaluated using PSNR, SSIM, and normalized correlation metrics, as well as its resistance to common image processing attacks. Particular attention is paid to analyzing the impact of digital watermarking on the quality of training data for computer vision models used in object classification and detection tasks. Experimental results demonstrate that watermark incorporation does not compromise visual data quality or reduce machine learning model accuracy, thereby confirming the feasibility of using QR decomposition for practical dataset protection