Abstract: Deepfake videos are a growing social issue. These videos are manipulated by artificial intelligence (AI) techniques (especially deep learning), an emerging societal issue. Malicious individuals misuse deepfake technologies to spread false information, such as fake images, videos, and audio. The development of convincing fake content threatens politics, security, and privacy. The majority of deepfake video detection methods are data-driven. This survey paper aims to thoroughly analyse deepfake video generation and detection. The paper’s main contribution is the classification of the many challenges encountered while detecting deepfake videos. The paper discusses data challenges such as unbalanced datasets and inadequate labelled training data. Training challenges include the need for many computational resources. It also addresses reliability challenges, including overconfidence in detection methods and emerging manipulation approaches. The research emphasises the dominance of deep learning-based methods in detecting deepfakes despite their computational efficiency and generalisation limitations. However, it also acknowledges the drawbacks of these approaches, such as their limited computing efficiency and generalisation. The research also critically evaluates deepfake datasets, emphasising the necessity for good-quality datasets to improve detection methods. The study also indicates major research gaps, guiding future deepfake detection research. This entails developing robust models for real-time detection.
The Evolving Threat of Deepfakes
As generative AI moves from niche research to user-friendly applications, the barrier to creating "perfect" fakes has vanished. This survey provides a comprehensive roadmap of the ongoing "arms race" between deepfake generators and detectors. While current deep learning models boast over 90% accuracy on standard benchmarks, they often crumble in real-world scenarios due to high computational costs, a lack of temporal consistency, and vulnerability to adversarial attacks.
The "detection gap" is a primary concern: most current forensic tools are designed for static images and fail to account for the temporal dimension of video—how a face moves over time. This challenge is exacerbated by the rapid evolution of the technology, transitioning from limited awareness in 2019 to a continuous technological duel between offenders and defenders today.
Taxonomy of Deepfake Technologies
To address the detection gap, it is essential to understand the breadth of manipulation techniques. Deepfakes are not limited to video; they span visual, text-based, and audio domains.
Within the visual domain, manipulations include face-swapping, lip-syncing, and attribute manipulation. The survey identifies three core pillars of failure in current research:
- Data Scarcity: A lack of high-quality, labeled deepfakes for training compared to the abundance of real videos.
- Computational Overload: The impracticality of detecting fakes at 30-60 frames per second in live feeds.
- The "Static Game" Fallacy: Detectors trained on specific GAN-based fakes often become obsolete when new techniques, such as Diffusion Models, emerge.
Methodologies in Detection
This systematic literature review analyzes 132 high-impact papers to build a taxonomy of detection challenges. Current methodologies include:
- Deep Learning (DL): The dominant force, utilizing CNNs for spatial features and RNNs/LSTMs or Transformers for temporal sequences.
- Frequency Domain Analysis: Identifying mathematical "artifacts" (e.g., Discrete Cosine Transform) invisible to the human eye.
- Biological Signals: A novel approach looking for human inconsistencies, such as irregular eye-blinking patterns or the absence of a pulse (photoplethysmography).
- Blockchain: Proposed for provenance—creating an unchangeable digital trail from the camera to the screen.
Dataset Evolution and the Benchmarking Paradox
Datasets have evolved through three generations, from small, low-quality sets like UADFV to the current gold standard, the DeepFake Detection Challenge (DFDC), which features over 100,000 videos with varied lighting and backgrounds. However, the field faces a "Benchmarking Paradox." In the 2020 DFDC, top models hit 82.56% accuracy on public sets but plummeted to 65.18% against "black box" hidden datasets. This suggests models are overfitting to specific dataset quirks rather than learning the essence of a "fake."
Key Results and Performance
Top-tier models like XceptionNet and EfficientNet remain the workhorses of the field, achieving 91% to 99% accuracy on datasets like FaceForensics++. MesoNet, a "shallow" CNN, offers a trade-off by providing higher speed for near-real-time detection, while Capsule Networks show high resilience to "replay attacks" (recording a fake off a screen). Despite these successes, as accuracy increases, computational requirements scale exponentially, hindering social media deployment.
Future Directions
The survey concludes that real-time detection remains largely a myth for high-accuracy models. Future research must pivot toward "Physics-guided AI"—combining deep learning with the laws of optics and human biology—and unsupervised learning (one-shot or few-shot learning) to identify new types of fakes without requiring massive labeled datasets. Furthermore, the lack of explainability in "Black Box" detectors remains a significant hurdle for legal and high-stakes moderation use cases.
Summary updated