A Survey of Defenses Against AI-Generated Visual Media: Detection,Disruption, and Authentication

TL;DR

This survey addresses defenses against AI-generated visual media by organizing detection (passive) and proactive defenses—disruption and authentication—into a unified generation-pipeline framework. It introduces multidimensional taxonomies for each defense strategy, surveys corresponding datasets and evaluation metrics, and analyzes trustworthiness concerns such as robustness and fairness. The work synthesizes advances across image and video forgery detection, fine-grained localization, attribution, and sequential manipulation, and reviews disruption and watermark-based authentication with a focus on robustness to real-world degradations and attacks. By outlining challenges and concrete directions, the paper aims to guide researchers and practitioners in building reliable, generalizable defenses and credible provenance tools for synthetic media.

Abstract

Deep generative models have demonstrated impressive performance in various computer vision applications, including image synthesis, video generation, and medical analysis. Despite their significant advancements, these models may be used for malicious purposes, such as misinformation, deception, and copyright violation. In this paper, we provide a systematic and timely review of research efforts on defenses against AI-generated visual media, covering detection, disruption, and authentication. We review existing methods and summarize the mainstream defense-related tasks within a unified passive and proactive framework. Moreover, we survey the derivative tasks concerning the trustworthiness of defenses, such as their robustness and fairness. For each defense strategy, we formulate its general pipeline and propose a multidimensional taxonomy applicable across defense tasks, based on methodological strategies. Additionally, we summarize the commonly used evaluation datasets, criteria, and metrics. Finally, by analyzing the reviewed studies, we provide insights into current research challenges and suggest possible directions for future research.

Figures (6)

• Figure 1: Framework of proactive and passive defenses against AI-generated visual media, which consists of three defense strategies: disruption, authentication, and detection. These three defense strategies can be employed independently or collaboratively. Within each independent defense strategy, we review related mainstream tasks and their trustworthiness research. In addition, we survey the research efforts on their joint defense.
• Figure 2: Distribution of literature across different defense strategies, including detailed statistics on the number of publications for the contained tasks and the defense associated trustworthiness: disruption (29 papers), authentication (26 papers), and detection (132 papers).
• Figure 3: Paper structure and overview of detection. Mainstream detection tasks ($\vcenter{}$) include (i) forgery detection (Section \ref{['sec:forgery_detection']}) and fine-grained forgery detection (e.g., (ii) passive forgery localization (Section \ref{['sec:passive_forgery_localization']}), (iii) forgery attribution (Section. \ref{['sec:forgery_attribution']}), (iv) sequential manipulation prediction (Section \ref{['sec:sequential_manipulation_prediction']})). Mainstream trustworthiness issues ($\vcenter{}$) of detection include (v) detection robustness (Section \ref{['sec:detection_robustness']}) and (vi) detection fairness (Section \ref{['sec:detection_fairness']}).
• Figure 4: Illustration of global and/or local spatial-temporal artifacts in video-level detection.
• Figure 5: Paper structure and overview of disruption. Disruption can be performed by attacking ($\vcenter{}$) different modules of generative models, such as the encoder, generator, and decoder, by adding perturbations ($\vcenter{}$) to the inputs or conditions. Mainstream disruption tasks ($\vcenter{}$) include (i) distorting disruption (Section \ref{['sec:distorting_disruption']}) and (ii) nullifying disruption (Section \ref{['sec:nullifying_disruption']}). The most studied disruption trustworthiness issue ($\vcenter{}$) is (iii) disruption robustness (Section \ref{['sec:disruption_robustness']}).