Deep Fake Detection: Enhancing Content Verification Using CNN-LSTM for Mitigating AI-Generated Video Threats

The emergence of Generative Adversarial Networks has dramatically transformed the landscape of synthetic media, enabling the creation of hyper-realistic deep fake videos that are increasingly difficult to distinguish from authentic content. This poses a growing threat, particularly across social media platforms like YouTube, Facebook, and TikTok, where misinformation can spread rapidly and influence public discourse. Traditional video analysis techniques are often inadequate in addressing the sophistication of such forgeries, highlighting the urgent need for robust detection systems. In this study, we present a deep fake detection model that effectively combines the spatial capabilities of a ResNeXt-based convolutional neural network with the temporal sensitivity of a Long Short-Term Memory network. The architecture utilizes 2048-dimensional feature vectors extracted from final pooling layers, with dropout applied to mitigate overfitting, and LSTM layers capturing sequential dependencies between video frames. By tuning the learning rate and extending the architecture with additional layers, our model demonstrated strong performance. Notably, it achieved an accuracy of 84.8%, with precision of 89.1% and F1 score of 84.5%, reflecting its reliability in distinguishing real from fake content. The system also forms the basis of a functional web prototype, offering potential real-world applications in digital forensics, content verification, and online safety. Ultimately, this research supports broader efforts to maintain information integrity and combat the misuse of generative AI technologies.