A low-cost method for reliable ownership identification of medical images using SVM and Lagrange duality

Abstract

Prevention of false positive and false negative errors is a major challenge for ownership identification and proof of ownership applications using digital image watermarking. Such errors are more critical with sensitive data, such as electronic patient records (EPRs) in medical image watermarking. A false positive error is a watermark detection error, which means that a watermark is detected in a media where there is no watermark. In contrast, a false negative error is an inability of the watermark detector to detect an embedded watermark in a watermarked image. These errors make ownership assessments unreliable, and the incorrect ownership identification of a patient’s record could result in failure of the correct diagnostics and treatments. To address this type of problem, a low-cost technique based on a support vector machine (SVM) and Lagrange duality was proposed to achieve reliable approximations for ownership identification in medical image watermarking without requiring the correction of attacked watermarked images. In this technique, the results of the ownership evaluation are categorized into two independent classes, namely watermark-detected and watermark-not-detected, and higher geometric margins between these classes are associated with higher reliability. To address additional situations with false positive and false negative errors, four different situations, including watermarked, unwatermarked, attacked watermarked and attacked unwatermarked images, were investigated. Experiments were conducted on duo-ISB-bit-plane (BiISB) watermarking using the histogram intersection (HI) technique as a testing platform under JPEG2000 and JPEG image compression attacks and using two groups of images: standard image processing images and X-ray medical images. The experimental investigations revealed that the HI technique guarantees that the rightful owner can be reliably identified even after severe attacks and in the face of context similarities between the watermark and the embedding pixels of the host image.