Quantifying the Impact of Watermarking on Deep Learning Accuracy in Medical Image Classification

Medical imaging is a crucial component of any modern healthcare system, and image classification is a key technique of Computer-Aided Diagnosis (CADx) system that provides accurate detection and treatment for various diseases. With the increasing use of digital medical images, there is a growing concern for their security and protection against unauthorized access, tampering, and copying. However, the impact of watermarking on the accuracy of image classification models, particularly in medical imaging, remains an undiscovered area of research. To the best of our knowledge, this work is one of the pioneering attempts in the field that examines the impact of watermarking on image classification accuracy with deep learning. Within this context, brain tumor and histopathologic cancer datasets are utilized. A Light-weight Conventional Neural Netwok (CNN) is developed, and several CNN models, such as ResNet-50, VGG16, VGG19, DarkNet-19, and AlexNet, are specially tailored to achieve accurate classification. Following the evaluation of the classification performance, a new watermarking scheme based on the Redundant Discrete Wavelet transform (RDWT), Multiresolution Singular Value Decomposition (MSVD), and Arnold Cat Map (ACM) is applied to the images, the classification accuracy of the watermarked testing-sets is subsequently re-evaluated. Results indicated that the proposed watermarking scheme is highly robust as the Normalized Correlation (NC) exceeded 0.9 and up to 1 for some attacks, and imperceptibility is well maintained at optimal levels as the Peak-Signal-to-Noise Ratio (PSNR) ranged from 66 dB and up to a maximum value of nearly 68 dB. Additionally, analyzing the effect of watermarking on classification accuracy indicated that the effect is found to be minimal, and ranged from 0% to 0.15% at most.