<![CDATA[The rapid advancement of generative artificial intelligence enables the creation of highly realistic fake videos, escalating the risks of financial fraud, hoaxes, and cybersecurity threats. As conventional visual verification methods are increasingly inadequate against evolving manipulation techniques, this research proposes a robust and lightweight AI-generated video detection approach. This method combines Local Binary Pattern (LBP) for spatial texture extraction with a Support Vector Machine (SVM) for binary classification. Utilizing the SDFVD2.0 dataset of 927 video samples, the methodology extracts frames and applies rigorous preprocessing, including grayscale conversion, resizing, noise reduction, and face cropping. To capture local micro-texture characteristics, LBP features are aggregated across frames using statistical mean and standard deviation, accounting for temporal dynamics and forming a comprehensive 144-dimensional feature vector. Subsequently, a linear SVM, optimized with balanced class weights and a soft-margin penalty, classifies these vectors. The LBP-SVM model achieved an accuracy of 81.18% on the testing split. During further generalization testing on unseen data, the model correctly predicted three out of five videos with a 62.72% average confidence rate and a swift processing time of 38.13 seconds. Although the model shows margins of error with highly complex artifacts, this combination provides an efficient, interpretable, and computationally economical baseline.]]>
