Shulou(Shulou.com)11/24 Report--

CTOnews.com, August 2, Google Research team is conducting an experiment in which they use OpenAI's GPT-4 to break the security measures of other AI models. The team has so far breached the AI-Guardian audit system and shared relevant technical details.

After inquiry, CTOnews.com learned that AI-Guardian is an AI audit system that can detect whether there is improper content in the picture, and whether the picture itself has been modified by other AI. If it detects the above signs in the picture, it will prompt the administrator to deal with it.

In a paper entitled "LLM-assisted Development of AI-Guardian", Nicholas Carlini, a researcher at Google Deep Mind, discussed the use of GPT-4 to "design attack methods and write attack principles", and used these schemes to deceive the defense mechanism of AI-Guardian.

The ▲ source Google research team has learned that GPT-4 will send a series of wrong scripts and explanations to deceive AI-Guardian. According to the paper, GPT-4 can make AI-Guardian think that "a picture of someone with a gun" is a picture of "someone with a harmless apple", thus allowing AI-Guardian to release the relevant image input source directly. The Google research team said that with the help of GPT-4, they had successfully "cracked" AI-Guardian 's defenses, reducing the accuracy of the model from 98 per cent to just 8 per cent.

At present, the relevant technical documents have been published in ArXiv, interested friends can go to learn about it, but the developers of AI-Guardian also pointed out that this attack method of the Google research team will no longer be available in future AI-Guardian versions, considering that other models will follow, so the current Google attack plan can only be used for reference in the future.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.