TL;DR
This survey provides a comprehensive, unified framework for defending against AI-generated visual media, categorizing strategies into detection, disruption, and authentication. It moves beyond simple "fake vs. real" classification to address fine-grained analysis and the critical, often-overlooked issues of defense trustworthiness, such as robustness and fairness.
What Problem Does This Paper Address?
The rapid advancement of deep generative models—like GANs and Diffusion Models—has made it trivial to create high-fidelity "Deepfakes" and synthetic scenes. While these tools have creative uses, they also enable misinformation, identity theft, and copyright infringement.
The research community has responded with various defense mechanisms, but the landscape is fragmented. Most existing surveys focus solely on passive detection (identifying a fake after it's released) and ignore proactive measures or the "trustworthiness" of the defenses themselves (e.g., can a detector be easily fooled by noise, or is it biased against certain demographics?). This paper aims to unify these disparate efforts into a single, systematic taxonomy.
Methods
The authors propose a unified framework that follows the standard AI generation pipeline, categorizing defenses into three main strategies:
- Detection (Passive): Identifying fakes after they are released. The paper categorizes these based on the artifacts they target (e.g., physical inconsistencies like eye blinking, or frequency-domain traces from up-sampling) and the learning strategies used (e.g., self-supervised learning or multi-task learning).
- Disruption (Proactive): Interfering with the generation process before it happens. This involves adding "adversarial perturbations"—tiny, invisible bits of noise—to real images so that if a generative model tries to manipulate them, the output becomes distorted or useless.
- Authentication (Proactive): Injecting digital watermarks into media. These can be robust (surviving the generation process to prove ownership) or semi-fragile (breaking when the image is tampered with to highlight exactly where a forgery occurred).

Crucially, the paper also surveys Trustworthiness, focusing on Robustness (resistance to adversarial attacks) and Fairness (ensuring detectors perform equally well across different races, ages, and genders).
Datasets
The survey reviews a vast array of datasets, categorizing them into two main types:
- Deepfake Datasets: Focused on facial manipulations (e.g., FaceForensics++, Celeb-DF, and the massive Deepfake Detection Challenge (DFDC)).
- Universal Generative Datasets: Aimed at detecting any AI-generated content, not just faces (e.g., CNNDetection, GenImage, and GenVideo).
The authors note a trend toward larger, more diverse datasets that include "in-the-wild" samples and multi-modal data (audio + video) to better reflect real-world challenges.
Key Insights and Ideas
- The Dominance of Detection: As shown in Figure 2, detection remains the most heavily researched area (71% of the literature), while proactive measures like disruption and authentication are still emerging.
- From Hand-crafted to Deep Features: Early defenses looked for specific "tells" like missing reflections in eyes. Modern methods use deep learning to find subtle, high-dimensional artifacts that are invisible to humans but distinct to specific model architectures.
- The Generalization Gap: A major "aha moment" in the field was realizing that a detector trained on GAN-generated images often fails completely on Diffusion-generated images. This has shifted research toward "universal" detectors that look for fundamental traces of the generation process itself.
- Trustworthiness is the New Frontier: A defense is only useful if it's reliable. The paper highlights that many top-tier detectors are vulnerable to "detector-agnostic" attacks that remove forgery traces without changing the image's appearance.

Key Results
The paper provides extensive lookup tables (Tables 7-10) comparing state-of-the-art methods.
- Detection: Top methods on benchmarks like FaceForensics++ achieve near-perfect AUCs (>99%) in-domain, but performance can drop significantly (sometimes to 60-70%) when tested on different datasets, highlighting the generalization problem.
- Disruption: Methods like PhotoGuard and Anti-DreamBooth show that adding minimal noise can effectively "protect" images from being used to train personalized generative models.
- Authentication: Modern watermarking schemes like Stable Signature can achieve 99% accuracy in verifying the origin of an image while maintaining high visual quality (PSNR > 30dB).
Limitations and Future Work
The authors identify several critical gaps:
- Fine-grained Interpretation: We need more than just a "fake" label; we need to know what was changed and which model did it (forgery attribution).
- Universal Benchmarking: While Deepfake detection has standard benchmarks, "universal" media detection (detecting any AI image) lacks authoritative, widely recognized evaluation sets.
- Practicality Metrics: Most research focuses on accuracy, but in the real world, inference speed and model size are just as important for deploying defenses on social media platforms.
- Fairness in Proactive Defense: While fairness is being studied in detection, it is almost entirely ignored in disruption and authentication research.
Summary updated