GenAI Safety

Generative AI systems, particularly diffusion-based text-to-image (T2I) models, are rapidly transforming creative and assistive technologies. These models align natural language with highly realistic image generation and are already used by millions worldwide. However, this tight coupling of language and imagery introduces unforeseen risks: sensitive attributes embedded in text—such as personal writing style, identity markers, or even pathological speech characteristics—can inadvertently propagate into generated images.

Our research reveals that adversaries can infer authorship and even detect medical conditions such as dementia not only from textual prompts but also from the images alone. We rigorously demonstrate that pathological speech markers and stylistic attributes survive the diffusion process and leak into the visual domain, with inference accuracy reaching state-of-the-art levels. This leakage poses profound privacy threats, enabling unauthorized profiling, re-identification, or discrimination of vulnerable users. To explain and quantify these risks, we design novel black-box adversarial models, text–image embedding analyses, and explainability-driven metrics that expose how linguistic information transforms and persists in generative outputs. In response, our research spearheads the design of privacy-preserving and accountable safeguards for the ethical and responsible deployment of safe generative AI models.

Recent Publications: