Caixin News, May 12th (Editor Niu Zhanlin) - The UK Artificial Intelligence (AI) Security Research Institute released a new testing platform on Friday aimed at strengthening the monitoring of security risks in advanced AI models.
It is reported that the toolbox is called Inspect and can be used to evaluate AI models in a range of fields, including their core knowledge, reasoning ability, and autonomy. By releasing under an open source license, this means that Inspect can be used for free by the global AI community.
Last October, the UK announced the establishment of the Artificial Intelligence Security Research Institute, which will research and test new AI models; In February of this year, the UK also announced that it will invest over £ 100 million to launch nine new research centers and provide technical training to AI regulatory agencies.
At a press conference, the UK Institute for Artificial Intelligence Security stated that Inspect is a software library that allows testers to evaluate the specific capabilities of individual AI models and then give a score based on the results.
Inspect can be used starting from Friday, which is also the AI security testing platform first launched by nationally supported institutions.
Under the current wave of AI competition, more and more AI models will be launched this year, making it more urgent than ever to promote the development of AI security.
However, it is still quite difficult to benchmark AI models at present, as the most complex AI models today are basically "black boxes", and their infrastructure, training data, and other key details are usually kept confidential by the companies that created them and not disclosed to the public.
So, how did Inspect address this challenge? Mainly through its scalability, it can adapt and accept new testing technologies. The built-in components of Inspect can be enhanced or extended through third-party software packages written in Python.
Inspect consists of three basic components: dataset, solver, and scorer. The dataset is used to evaluate the sample set for testing, the solver is a component that performs actual testing work, and the scorer is responsible for evaluating the work results of the solver, ultimately generating a comprehensive evaluation of the performance of the AI model. This design allows Inspect to flexibly adapt to different testing needs and evaluation standards.
As part of the UK's ongoing leadership in the field of AI security, I have approved the open source Inspect, which demonstrates the UK's unique talent and creativity in innovation and technological development, and consolidates our position as a world leader in this field, according to UK Minister of Science Michelle Doneland.
Ian Hogarth, Chairman of the Institute of Artificial Intelligence Security, claims that successful AI security testing collaboration means having a shared and accessible evaluation method, and we hope that Inspect can become the cornerstone of AI security research, research organizations, and academia.