Security and Privacy on Generative Data in AIGC: A Survey

TL;DR

This survey systematically review the security and privacy on generative data in AIGC, particularly for the first time analyzing them from the perspective of information security properties, and reveals the successful experiences of state-of-the-art countermeasures in terms of the foundational properties of privacy, controllability, authenticity, and compliance.

Abstract

The advent of artificial intelligence-generated content (AIGC) represents a pivotal moment in the evolution of information technology. With AIGC, it can be effortless to generate high-quality data that is challenging for the public to distinguish. Nevertheless, the proliferation of generative data across cyberspace brings security and privacy issues, including privacy leakages of individuals and media forgery for fraudulent purposes. Consequently, both academia and industry begin to emphasize the trustworthiness of generative data, successively providing a series of countermeasures for security and privacy. In this survey, we systematically review the security and privacy on generative data in AIGC, particularly for the first time analyzing them from the perspective of information security properties. Specifically, we reveal the successful experiences of state-of-the-art countermeasures in terms of the foundational properties of privacy, controllability, authenticity, and compliance, respectively. Finally, we show some representative benchmarks, present a statistical analysis, and summarize the potential exploration directions from each of theses properties.

Figures (10)

• Figure 1: The process of AIGC. Real data collected is used to train generative models. Then generative models produce generative data. Finally, generative data are further analyzed. For generative data, there are corresponding protection requirements of security and privacy at different stages, which can be divided into privacy, controllability, authenticity, and compliance.
• Figure 2: The subclassification of security and privacy on generative data.
• Figure 3: The example of privacy in AIGC is from carlini2023extracting, and the example of AIGC for privacy is from chen2021perceptual.
• Figure 4: Generated samples by Stable Diffusion just replicate training data via piecing together foreground and background objects in training data somepalli2023diffusion.
• Figure 5: Examples of different protections for malicious access from yeh2021attack, LYSBFZ23, and 10086559.