A SCALABLE HYBRID ARCHITECTURE FOR REAL-WORLD DEEPFAKE DETECTION AND VERIFICATION

Abstract

Artificial intelligence has grown in very fast speeds that have created hyper-realistic deepfakes through Generative Adversarial Networks (GANs) and diffusion models that jeopardize the integrity of information, personal identity, and international trust. Modern detection systems frequently have problems with generalization and compression sensitivity, being unable to function when dealing with new methods of manipulation or real-world social media images. To address these shortcomings, this work presents a proposal of a scalable hybrid architecture that will connect multi-dimensional signals into one coherent verification pipeline. In accordance with the CRISP-DM approach, the framework incorporates three main modules, namely, a Psychological and Behavioral one, which studies the micro-expressions, a Physiological Sensing one, which identifies the subconscious indicators such as pulse rate (rPPG), and a Visual Artifacts one. The architecture is based on a series of 10 stages with a lightweight MobileNetV3 backbone and Graph Convolutional Networks (GCN) to predict structural relationships among regions of the image. It is followed by the use of a four-head Transformer block to capture long-range dependencies and global context. The system is trained on a variety of datasets, such as FaceForensics++, Celeb-DF, and ForgeryNet, so that the system will still be resistant to cross-dataset perturbations.

. Initial tests have shown a detection accuracy of about 95, AUC scores of over 0.90 and verification time of less than seven minutes. This lightweight design, a defense-in-depth strategy that reduces the liar’s dividend (to less than 8%), provides a robust solution to a wide range of digital realities, such as personal identity protection or large-scale journalistic verification.