{"id":88716,"date":"2024-10-02T17:59:01","date_gmt":"2024-10-02T15:59:01","guid":{"rendered":"https:\/\/intercoaching.fr\/?p=88716"},"modified":"2024-10-02T19:36:36","modified_gmt":"2024-10-02T17:36:36","slug":"nightshade-the-formidable-weapon-to-thwart-text-to-image-ai-models","status":"publish","type":"post","link":"https:\/\/intercoaching.fr\/en\/nightshade-the-formidable-weapon-to-thwart-text-to-image-ai-models\/","title":{"rendered":"Nightshade, the formidable weapon to thwart text to image AI models"},"content":{"rendered":"
Nightshade: the intrusion system into text image AI<\/h2>\n\n\n
Computer science specialists from the University of Chicago have just developed a tool, called Nightshade, which has the capacity to disrupt the automatic learning phase of AIs oriented towards generating images from texts. This approach aims to introduce altered data into these intelligent models.<\/p>\n\n\n
The discreet subversion of image banks<\/h3>\n\n\n
Artificial intelligence dedicated to converting text into images requires large, high-quality databases, often made from eclectic collections available on the Internet. However, the harvesting of this massive data, sometimes in violation of copyright laws, is a common practice among the designers of such systems, including well-known solutions such as Dall-E 3 or Stable Diffusion.<\/p>\n\n\n
A stealthy but impactful manipulation<\/h2>\n\n\n\n\n<\/figure>\n\n\n
With the aim of protecting the rights of artists, researchers at the University of Chicago therefore developed Nightshade, a mechanism that imperceptibly subverts the content of images. Although the transformations made on the images are undetectable by the human eye, they induce errors in the interpretation of the AI \u200b\u200bmodels, which leads to disruptions during their training. An illustrative case would be an image of a wolf which, once Nightshade\u2019s manipulations were applied, would be perceived by the algorithm as an image of a car during its machine learning, thus causing inadequate responses from the AI.<\/p>\n\n\n
\n\n
The image labels for the model are distorted.<\/li>\n\n\n
The tool is especially effective against Stable Diffusion algorithms.<\/li>\n\n<\/ul>\n\n\n
Defend the integrity of models against corruption<\/h2>\n\n\n\n\n<\/figure>\n\n\n
Nightshade poses a real threat to the reliability of the affected models, since it is enough to introduce a small percentage of corrupted images into the training stream to significantly alter their capabilities. Disconcerted by the sophistication of the tool, experts have not yet found foolproof solutions to thwart these alterations, but are proposing reaction mechanisms, such as the elimination of problematic images during learning or updating. point of a detection system in response to Nightshade, plunging developers into a constant struggle for the security of their models.<\/p>\n\n\n\n
![\"nightshade,](\"https:\/\/intercoaching.fr\/wp-content\/uploads\/2023\/12\/Nightshade-larme-redoutable-pour-dejouer-les-modeles-dIA-text-to-image.png\")
\n<\/figure>\n\n\n![\"nightshade,](\"https:\/\/intercoaching.fr\/wp-content\/uploads\/2023\/12\/Nightshade-larme-redoutable-pour-dejouer-les-modeles-dIA-text-to-image-1.png\")
\n<\/figure>\n\n\n