English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Text-to-image models like Stable Diffusion have had a profound impact ondaily life by enabling the generation of photorealistic images from textualprompts, fostering creativity, and enhancing visual experiences across variousapplications. However, these models also pose risks. Previous studies havesuccessfully demonstrated that manipulated prompts can elicit text-to-imagemodels to generate unsafe images, e.g., hateful meme variants. Yet, thesestudies only unleash the harmful power of text-to-image models in a passivemanner. In this work, we focus on the proactive generation of unsafe imagesusing targeted benign prompts via poisoning attacks. We propose two poisoningattacks: a basic attack and a utility-preserving attack. We qualitatively andquantitatively evaluate the proposed attacks using four representative hatefulmemes and multiple query prompts. Experimental results indicate thattext-to-image models are vulnerable to the basic attack even with fivepoisoning samples. However, the poisoning effect can inadvertently spread tonon-targeted prompts, leading to undesirable side effects. Root cause analysisidentifies conceptual similarity as an important contributing factor to theside effects. To address this, we introduce the utility-preserving attack as aviable mitigation strategy to maintain the attack stealthiness, while ensuringdecent attack performance. Our findings underscore the potential risks ofadopting text-to-image models in real-world scenarios, calling for futureresearch and safety measures in this space.

On the Proactive Generation of Unsafe Images From Text-To-Image Models Using Benign Prompts