Enhancing Deepfake Detection: A Weighted Summation Model of CNN Approach with Local and Global Analysis

In recent years, the advent of sophisticated technologies capable of producing highly realistic images has blurred the line between authenticity and forgery, posing significant societal challenges. The human eye often struggles to discern genuine images from their artificially generated counterparts, necessitating reliable methods to ascertain image authenticity. Scholars have strived to enhance detection precision by refining Convolutional Neural Network (CNN) models. Despite numerous CNN-based models achieving detection accuracies of up to 90%, they are not without their flaws, including limited generalisability, suboptimal parameter solidification, and inadequate scenario-specific refinement, alongside insufficient preprocessing of image data. To address these shortcomings, this research proposes a novel CNN model as a foundation. The image dataset is initially segmented into four equal parts, upon which five independent CNN models are trained, both on these subdivisions and the image in its entirety. The incorporation of a Bayesian model facilitates automatic parameter tuning, followed by a weighted summation of the activation function values derived from the quintet of models, each optimised through training. The aggregated outcome is then evaluated against pre-established criteria to ascertain the veracity of images under various conditions.