Interpretable AI-Generated Videos Detection using Deep Learning and Integrated Gradients

The rapid advancements in generative AI have led to text-to-video models creating highly realistic content, raising serious concerns about misinformation spread through synthetic videos. As these AI videos become more convincing, they threaten information integrity across social media, news, and digital communications. Using AI-generated videos, bad actors can now create false narratives, manipulate public opinion, and influence critical processes like elections. This technology's democratization means that sophisticated disinformation campaigns are no longer limited to well-resourced actors, creating an urgent need for reliable detection methods and human-machine cooperation to maintain public trust in visual information across our digital transformation landscape. The accessibility of these tools to a broader audience amplifies the potential for widespread misinformation, making robust detection systems crucial for maintaining social media integrity.Through our research into video generation models, we identified that state-of-the-art systems like diffusion transformers operate on patches of noisy latent spaces. We deliberately mirrored this architecture in our classifier design, enabling it to analyze videos using the same fundamental structural unit generation models used to create them. This architectural alignment allows our system to adapt to emerging generation techniques while maintaining detection efficacy.We designed an explainable video classifier using deep learning and neural networks that detect AI-generated content and show evidence for its decisions. The classifier uses three main parts: a convolutional encoder that turns video frames into latent representations, a patch vectorizer that breaks these representations into analyzable chunks, and a transformer that processes these chunks to make the final decision. This human-centered computing design lets us efficiently process videos while maintaining explainability through Integrated Gradients, which reveal which input parts influenced the model's decisions.We use integrated gradients to show which parts of a video led to the model's decision. This method looks at how the model's decision changes as we move from a blank video to the actual video, showing us which pixels matter most for classification. These pixel-level maps provide clear evidence of why the model thinks a video is AI-generated or real, providing transparency critical for building trust in automated content verification systems.We will test our model on the GenVideo dataset, a comprehensive collection of videos labeled as real or AI-generated from diverse sources, including Stable Diffusion, Sora, Kinetic 400, and MSRVTT. This large-scale data analytics evaluation will check how well it classifies videos and explains its decisions, helping determine if the model can work as a practical tool for machine learning-based content verification, considering that wrong AI classifications could harm content creators' reputations.Our work adds to the growing field of explainable AI in content authentication and shows why we need clear evidence when making high-stakes decisions about video content. Future work will look at detecting hybrid videos (real videos with AI elements added) and making our visual explanations more useful for human decision-makers in content verification. The insights gained will inform the development of more sophisticated detection systems capable of addressing evolving challenges in digital content authenticity.