Author: Dona Roy
Affiliation: Department of Computer Science, National Institute of Technology, Karnataka
Email: donaroy.242is011@nitk.edu.in
Date: May 24, 2025
The rapid evolution of artificial intelligence and deep learning has significantly enhanced the realism of deepfakes, challenging media authenticity and public trust. This literature survey critically examines detection methodologies, datasets, and architectural advancements from 2017 to 2025, while highlighting ethical and legal implications.
_Fig 1: Deepfake : A Hyper-Realistic fake image.
- ✔️ Survey of 29+ deepfake detection models (CNNs, Transformers, Hybrid models)
- 🎯 Detection accuracy up to 92% (CNN-NeRF) and 89% (DiT-based Transformers)
- 🧩 Analysis of 11+ benchmark datasets including SynthID, DeepReal, and FaceForensics++
⚠️ Covers real-world misuse cases, including scams and political misinformation- 🔐 Emphasis on ethical, legal, and privacy challenges
| Dataset | Year | Volume | Techniques | Challenges |
|---|---|---|---|---|
| FaceForensics++ | 2019 | 1.8M images | FaceSwap, DeepFakes | Demographic bias, outdated |
| DeepReal | 2024 | 80K clips | DiT, NeRFace | High computational demand |
| SynthID | 2023 | 50K images | StyleGAN3, Diffusion | Excludes NeRF-based fakes |
| WildDeepfake | 2021 | 707 videos | Internet-sourced | Consent and labeling issues |
🔗 Dataset Links
| Category | Example Features |
|---|---|
| Visual Artifacts | Lip mismatch, facial asymmetry |
| Biological Artifacts | Skin/hair aging, blinking abnormalities |
| Frequency Domain | High-frequency noise in diffusion models |
| Spatio-Temporal | Frame coherence, unnatural lighting |
| Latent/Deep Features | GAN fingerprints, NeRF inconsistency |
- ResNet18 + ELA
- VGG-19 + CapsNet
- Gram-Net
- TruFor with Contrastive Noiseprint Learning
- Vision Transformer (ViT)
- DiT (Diffusion Transformer)
- Swin Transformer
- Self-Subtracting Transformer (with explainability)
- CNN + NeRF 3D feature modeling
- GAN-inverted artifact extraction
- Diffusion model-based detection (Stable Diffusion, DALL-E 2)
⚖️ Key Legal Cases
- 💼 2019: €220K corporate scam with voice deepfake
- 🎤 2024: $25M fraud with VALL-E clone
- 📹 2025: UN deepfake disrupted climate negotiations
- Identity theft
- Misinformation in politics
- Non-consensual explicit content
📌 Emphasis on dataset consent, privacy preservation, and responsible AI policies.
✅ Real-time detection pipelines
✅ Multi-modal analysis (audio, video, physiological)
✅ Brain-inspired learning (SNNs, neuromorphic architectures)
✅ Quantum-enhanced feature extraction
✅ Legally enforceable digital watermarking & provenance (e.g., blockchain)
See full Reference List in the main paper.
@misc{roy2025deepfake,
title={Deepfake Detection: A Literature Review},
author={Dona Roy},
year={2025},
institution={NIT Karnataka},
note={Accessed from personal survey}
}