I need a lightweight AI model that can flag deepfakes in both still photos and the individual frames of a video. The goal is moderate, reliable accuracy—good enough to separate obvious fakes from genuine content without the heavy-duty infrastructure a research-grade solution would demand. I just want clean, well-commented code that someone with basic ML/DL familiarity can pick up and run on a CPU. Deliverables • Trained model file and all training scripts • Inference script or notebook that accepts an image or video file and returns a deepfake probability score plus a simple real/fake label • Brief README explaining setup, model architecture, and how to retrain or fine-tune on new data • A short report (a page or two) summarizing achieved accuracy, dataset used, and suggested next steps for further improvement If public datasets are required, please source ones that are freely licensed and reference them in the report. Keep the solution straightforward—clear code, minimal external dependencies, and results that demonstrate solid, good-accuracy detection without an oversized model.