Exploiting Visual Artefact from Deepfake Videos Using Hybrid Meta-heuristic Optimisation Algorithm

The widespread flow of deepfake videos across social media platforms has raised serious security concerns. Despite significant progress in traditional deepfake detection methods, several challenges persist, such as distortion from compression artefacts, loss of critical feature information, high execution time, suboptimal feature selection, poor handling of dynamic or multi-face content, data imbalance, limited labelled training data, and disruptions in feature distribution, which continue to degrade detection accuracy. To address these limitations, this paper presents a Secretary Bird Skill Optimisation Algorithm-based Pyramid Deep Belief Network (SB-PyDBNet) for robust multi-face deepfake recognition. Initially, frames from the input video are extracted. After that, the face is detected employing YOLO v3-Tiny and next, facial action unit detection is accomplished by Action Unit Network (AUNet). Thereafter, ResNet features, statistical features and Histogram of Oriented Gradients (HOG) are extracted. Finally, multi-face deepfake recognition is done utilising Pyramid Deep Belief Network (PyramidFDBNet), which is introduced by combining Deep Pyramidal Residual Network (PyramidNet) with Deep Belief Network (DBN), whereupon modification of layers is performed by the Fractional Calculus (FC) concept. Moreover, PyramidNet training is performed by the Secretary Bird Skill Optimisation Algorithm (SBSOA), which is devised by integrating the Secretary Bird Optimisation Algorithm (SBOA) with the Skill Optimisation Algorithm (SOA). Additionally, SBSOA_PyramidFDBNet acquired 92.378% accuracy, 93.790% True Negative Rate (TNR), and 91.194% True Positive Rate (TPR).