Deepfake Video and Image Detection using Deep Learning

Detecting synthetic media is critical to safeguarding digital ecosystems against misinformation and identity abuse. We present a two-track deep learning system for image and video deepfake detection. For images, we build a lightweight convolutional neural network that operates on 224×224 RGB inputs with standard normalization and trains end-to-end using categorical cross-entropy. On a two-class dataset (“Real”, “Fake”, 10,000 images), the model attains 95% accuracy with balanced precision/recall (Fake: 0.96 F1; Real: 0.95 F1). For videos, we model temporal cues by combining per-frame spatial features with recurrent aggregation. Faces are detected and cropped from frames; each crop is resized to 128×128 and passed through a compact CNN feature extractor (32-D embedding). Sequences of 30 consecutive embedding are then classified by a two-layer LSTM followed by a sigmoid output. Evaluated on 5,5,548 labelled sequences, the video pipeline achieves 98% accuracy (Real: 0.88 F1; Fake: 0.88 F1), indicating robust performance on spatio-temporal artifacts. We provide saved models and an inference routine that consumes raw videos and returns calibrated real/fake predictions. Our results show that (1) even compact convolutional backbones yield strong image-level performance, and (2) explicit temporal modelling substantially improves video detection compared to per-frame decisions. We discuss deployment considerations, including label consistency, identity-disjoint splits, and stronger face detection/tracking, and outline future work with transfer learning backbones and frequency-domain cues to further harden the system against next-generation generative models.