<![CDATA[The rapid advancement of generative artificial intelligence has substantially improved the visual quality of synthetic video, making it increasingly difficult to distinguish from authentic footage and creating new challenges for digital content verification. This study develops and evaluates an AI-generated video detection system using a hybrid approach that integrates a Convolutional Neural Network (CNN) as a feature extractor with Extreme Gradient Boosting (XGBoost) as the classifier. The research uses a 70,000-video subset of the GenVidBench-143k benchmark, selected through stratified sampling for computational efficiency. Feature extraction is performed in parallel using ResNet50, which captures high-level semantic features, and VGG16, which captures local texture detail; the resulting 2,048- and 512-dimensional vectors are concatenated into a single 2,560-dimensional representation per frame. A temporal aggregation stage transforms frame-level features into video-level features using Average Pooling and Max Pooling, and a Cost-Sensitive Learning strategy is applied through the scale_pos_weight, sample_weight, and max_delta_step parameters of XGBoost to address the marked class imbalance between real and AI-generated videos. The experimental results show that the hybrid ResNet50-VGG16 architecture combined with Average Pooling aggregation and scale_pos_weight optimization yields the best performance, achieving 92.93% accuracy, with 0.89 precision and 0.85 recall for the real-video class, and 0.94 precision and 0.96 recall for the AI-video class. A robustness analysis further shows that the model performs strongly on low-resolution video (<720p), reaching 94.00% accuracy, and maintains 88.00% accuracy on high-resolution video (>=720p); the modest decline at higher resolution reflects the increasingly seamless visual quality of modern generative models. These findings indicate that the hybrid CNN-XGBoost approach is a promising and computationally efficient strategy for synthetic video detection, although broader training-data variation is still required to reduce detection errors on high-quality video in real-world scenarios.]]>
