Introduction:
The emergence of trained models and AI-generated tools has ushered in a new era of content creation, responding dynamically to human-provided queries. However, this monumental transformation has sparked a global discourse. Lingering uncertainties regarding the faithfulness of AI-generated content have emerged after years of dedicated research. This begs the question of whether AI’s inability to discern right from wrong might lead it astray from its intended trajectory. Unlike humans, who possess the capacity to recognize and rectify errors, AI treats all content as a facet of human perspective, applying it wherever it finds utility. This raises a pivotal concern: what implications linger for human privacy if AI exploits personal data without explicit consent? The rapid evolution of AI introduces a plethora of risks that warrant careful consideration.
Keywords:
AIGC dispute, Privacy leakage in foundation models, Privacy leakage in generative models.
Literature Review:
AI systems and tools such as ChatGPT being capable of producing content that is virtually indistinguishable from human-created output, there is an increasing worry about the ethical utilization of this technology. It is observed from past studies, there is currently a noticeable absence of a shared framework and standardized terminology for establishing and documenting the responsible application of AI in the realm of content generation[1].Jan Philip Wahle : Proposed a 3 dimensional model and AI usage card to secure to research data from breaches.[1]. Over the past few years, there have been significant strides and remarkable developments in the field of generative modeling. OpenAI’s DALL·E [Ramesh et al., 2021] emerged as one of the first text-to-image models that garnered extensive public recognition.Its successor, DALL·E 2 [Ramesh et al., 2022], which also generates more complex and realistic images, was unveiled in April 2022, followed by Stable Diffusion. Following these AI trends Google also stepped forward, presented two text-to-image models that can generate photorealistic images: the diffusion-based model Imagen , and the Pathways Autoregressive Text-to-Image model (Parti) .
Diffusion Models can also use to make image-to-image but also generate videos from text, such as Runway [Runway, 2022], Make-A-Video [Singer et al., 2022], Imagen Video [Ho et al., 2022], and Phenaki [Villegas et al., 2022].[3]. Many artificial intelligence generated content (AIGC) models depend on text encoders that are trained using extensive internet data. This data might include social biases, toxicity, and other constraints that are naturally present in large language models. The problem of privacy vulnerabilities in GPT-2 highlights concerns about the model unintentionally reproducing sensitive information from its training data. This can lead to privacy breaches as generated text might contain fragments of private data. Later this problem modified in updated GPT-3 version but not such accurate. Users should be careful not to generate sensitive content with these models.
Continue…..