Distinguishing True and Fake Ultra-High Definition Images Using Relative DCT Analysis and Machine Learning
Amanpreet Singh Saimbhi · 2025
This paper addresses the challenge of distinguishing true Ultra-High Definition (UHD) images from fake ones, specifically those that are upscaled from lower resolutions, using various interpolation methods or Deep Neural Networks (DNNs). I propose a novel approach leveraging Discrete Cosine Transform (DCT) analysis to detect fake UHD content. Our method includes creating a reference image through downsampling and upscaling to compute relative DCT coefficients, which are used to analyze differences in energy spectrum between genuine and fabricated UHD images. I evaluate several techniques, including simple thresholding and Support Vector Machine (SVM) classifiers, to identify the authenticity of UHD images. Our experiments show that while simple thresholding methods with basic interpolation references perform well, they struggle with images generated by DNNs such as SRCNN and EDSR. I enhance our models by incorporating DNN-generated samples into the training set, significantly improving classification accuracy. The final SVM model achieves near-perfect accuracy in distinguishing real UHD images from fake ones, including those generated by advanced DNN methods. Our results demonstrate the effectiveness of combining DCT analysis with machine learning techniques for UHD image classification, offering a robust solution for detecting fake content in high-resolution media.