In a recent study published on arXiv, researchers from the University of California, Berkeley, have developed a new method for attacking GANs called Prompt-specific Poisoning (PSP). This technique involves injecting specific prompts into the input data that can manipulate the output generated by the model.
The PSP attack works by first identifying the most important words in a given prompt and then replacing them with synonyms or other related terms. The resulting modified prompt is then fed to the GAN, which generates an image based on this new input. By carefully selecting these prompts, the researchers were able to manipulate the output of popular text-to-image models such as DALL-E and Stable Diffusion.
In another study published in IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), researchers from Tsinghua University have developed a new method for attacking GANs called Query-free Adversarial Attack against Stable Diffusion. This technique involves injecting noise into the input data to create adversarial examples that can fool the model into generating incorrect output.
The researchers found that their attack was effective in manipulating popular image generation models such as DALL-E and Stable Diffusion, with success rates of up to 95%. They also demonstrated that this technique could be used to generate images that were not present in the original dataset, which has significant implications for privacy and security concerns.
In a third study published on arXiv, researchers from Carnegie Mellon University have developed a new method called Universal and Transferable Adversarial Attacks on Aligned Language Models. This technique involves injecting noise into the input data to create adversarial examples that can fool popular language models such as GPT-3 and BERT.
The researchers found that their attack was effective in manipulating these models, with success rates of up to 90%. They also demonstrated that this technique could be used to generate text that was not present in the original dataset, which has significant implications for privacy and security concerns.