Deepfake Detection

Detection of deepfakes—AI-synthesized or face-swapped images and videos—encompasses both visual forensics and neural network-based approaches. The problem domain includes identifying faces generated by GANs, detecting face swaps via facial reenactment, and distinguishing authentic from manipulated videos.

Key papers

• Exposing the Deception: Uncovering More Forgery Clues for Deepfake Detection — Information-theoretic framework extracting both local and global forgery clues via disentangled feature learning; uses mutual information losses to ensure comprehensive, orthogonal feature representations; achieves 0.983 AUC on FaceForensics++, 0.999 AUC on Celeb-DF-V2, 0.939 AUC on DFDC; strong cross-dataset generalization.
• Recurrent Convolutional Strategies for Face Manipulation Detection in Videos — Recurrent convolutional networks exploiting temporal discrepancies in frame sequences; uses landmark-based face alignment and bidirectional GRU for deepfake, Face2Face, and FaceSwap detection, achieving 96.9%, 94.35%, and 96.3% accuracy on FaceForensics++.
• FakeCatcher: Detection of Synthetic Portrait Videos using Biological Signals — Detection using photoplethysmography (PPG) signals; analyzes spatial coherence and temporal consistency of blood flow patterns to achieve 91%+ accuracy across multiple datasets; robust to generative model variations.
• MesoNet: A Compact Facial Video Forgery Detection Network — lightweight CNN architectures (Meso-4, MesoInception-4) for Deepfake and Face2Face detection at mesoscopic level; achieves 98% and 95% detection accuracy respectively; introduces first Deepfake dataset.
• In Ictu Oculi: Exposing AI Generated Fake Face Videos by Detecting Eye Blinking — physiological signal-based detection exploiting absence of natural eye blinking in synthesized faces; LRCN model achieving 0.99 AUC over 0.98 for CNN and 0.79 for hand-crafted baselines.
• Detecting GAN-generated Imagery using Color Cues — Forensic detection via GAN generator architecture analysis, identifying color channel overlap and suppressed saturation as distinguishing cues; achieves 0.7 AUC on fully synthetic images.
• Two-Stream Neural Networks for Tampered Face Detection — two-stream architecture for face swapping detection combining GoogLeNet classification (visual artifacts) with steganalysis-based triplet network (noise residuals); AUC 0.927 on SwapMe/FaceSwap dataset.
• Exposing Deep Fakes Using Inconsistent Head Poses — head pose inconsistency as forensic cue exploiting facial landmark mismatch in deepfake generation.
• Deepfake Detection: A Systematic Literature Review — comprehensive survey of 112 deepfake detection papers; systematic comparison of deep learning (77%), traditional ML (18%), and statistical approaches.
• FaceForensics++: Learning to Detect Manipulated Facial Images — largest facial forgery benchmark (1.8M+ images) with four manipulation methods; CNN-based detection achieves 99.26% accuracy.
• Deepfakes and Disinformation: Exploring the Impact of Synthetic Political Video on Deception, Uncertainty, and Trust in News — experimental evidence that deepfakes increase epistemic uncertainty and reduce trust in news media.

Related topics

• Media Forensics (broader field)
• Face Forensics (facial image manipulation)
• Synthetic media (generated content)
• Generative Models (GAN-based face synthesis)