Katarzyna Petru
Google has published a new security report detailing the increasing attempts to copy advanced artificial intelligence models. The targets of the attacks include Gemini, the company's flagship AI model. According to the document, attackers used over 100,000 queries to analyse and replicate the system's functioning. Google asserts that these actions are part of a broader wave of intellectual property theft attempts related to AI. The attacks are said to originate from countries such as North Korea, Russia, and China. However, the company emphasises that the threat primarily concerns model creators, not ordinary users.
What is “model extraction”?
Google defines these activities as so-called extraction attacks, which involve systematically probing a model using legitimate access. Practically, this means the mass generation of queries to understand the mechanisms of AI and using that knowledge to create a competitive system. Such operations may lead to the development of models in other languages or to replicating functionalities without access to the original training data.
Arms Race in the AI World
Experts from the Google Threat Intelligence Group warn that the company may merely be the first example of a larger problem. In recent years, competition in the AI sector has significantly accelerated, and tensions between technology companies are rising. New, advanced models from Asian firms have emerged on the market, boldly competing with leaders from the USA. Additionally, there are ongoing disputes regarding the use of others' technology to train one's own systems.
The Google report shows that with the development of artificial intelligence, the risk of technology theft is also increasing. Attacks involving mass analysis of models could become one of the main threats to companies developing AI. For users, this means one thing – the race for dominance in the world of AI is just getting started.
source: cnet.com
