I require a purpose-built deep learning model capable of reliably distinguishing authentic images from manipulated (deepfake) images. The scope is tightly focused on model design, training, and evaluation: developing an effective CNN-based architecture, training it on established deepfake datasets, and tuning it to perform robustly under real-world conditions. The implementation may use TensorFlow / Keras, PyTorch, or an equivalent framework, provided the entire training and inference pipeline is fully reproducible on a single modern GPU. I can supply standard datasets (e.g., FaceForensics++ image frames, DFDC samples) along with additional proprietary images if required. Please indicate if alternative datasets would materially improve performance. Deliverables • Well-documented source code with clear modular structure • requirements.txt (or equivalent) for environment reproduction • Trained model weights • A concise README detailing dataset preparation, training commands, and inference steps • Technical evaluation report including: Accuracy Precision / Recall F1-score ROC-AUC Confusion matrix on a held-out test split Acceptance Criteria • ≥ 90% balanced accuracy on the provided test dataset • Inference latency ≤ 50 ms per image on an RTX 3080–class GPU • Model generalizes beyond training data (no dataset leakage or overfitting artifacts) Additional Notes If you have prior experience in deepfake detection, face forgery analysis, or media forensics, please include a short note or reference to past work (paper, repository, or deployed system). This will help assess technical fit efficiently.