China Xintong Institute, Tsinghua University, and Ant Group jointly released the AI security testing platform "Ant Jian"

The large -scale application of artificial intelligence has given birth to more urgent requirements for AI security. On the afternoon of September 1st, at the "2022 World Artificial Intelligence Conference-Trusted AI Forum" in Shanghai, China Xintong Institute, Tsinghua University, and Ant Group jointly released the AI ​​security testing platform "Ant Jian" to provide AI model developers with a model for models. A one -stop evaluation solution with defense reinforcement to help developers one -click identification and mining model vulnerabilities to create more secure and reliable AI.

The reporter learned that this is also the industry's first AI security detection product for the entire data type of industrial scenarios. The platform is based on the risk confrontation practice of Ant for many years. It integrates more than 1,000 trusted AI patents and combines self -developed intelligent game confrontation technology. It can realize automated testing and quantitative scores of different data types such as text, images, tables, and sequences. And security enhancement.

Li Junkui, president of the Ant Group's Great Security Technology Department, introduced that the platform is embedded with the self -developed AI security computing engine, which has been improved by dozens of times. Production and automatic evolution attack means, "detection -defense -upgrade" for AI models can effectively discover the AI ​​model vulnerabilities and reduce systemic risks; in addition, the platform covers "random noise" and "migration attacks on the model of unknown models. "The attack samples are more in line with the attack evaluation benchmark of industrial applications.

Li Junkui said that the platform supports zero -coding evaluation. Users upload the model to be tested. After selecting the attack algorithm, the platform can automate the security risk of AI and comprehensively strengthen the model algorithm. The platform is now open to AI model robust evaluation, and in the future, it will continue to increase the detection services of model explanation, algorithm fairness, and privacy protection in the future.

According to reports, the technical capabilities equipped by the "Ant Jian" AI security detection platform have been effective application practice in Ant's trading security, nuclear body security, and content security. For example, in the spam control scenario, through the platform based on the deep neural network game offensive and defensive model, the simulation of black -producing attacks will be defensive in advance, and the risk account is missing by more than 20 %; Reduce risk leaks by 38%; in the voucher certification scenario, based on the model problems found in advance and the counter -training enhancement scheme, eliminate 23%of the potential black production attack threats.

Shi Lin, deputy director of the content and technology department of the Institute of Cloud Computing and Big Data of China Xintong Institute, said that artificial intelligence technology is deeply integrated in the fields of manufacturing, medical, energy, transportation, and urban construction, and the issue of AI algorithms is constantly appearing. The problems of algorithm black box, algorithm design defects, data dependencies, confrontation sample attacks and other issues bring great hidden security risks to the industrialized application of intelligent technology. AI is a large -scale application of industrial scenarios. It is necessary to achieve full -link security and credibility, and continue to strengthen model robustness.

Li Qi, an associate professor at Tsinghua University, said that the general automation testing of artificial intelligence algorithms is the core of research and solving artificial intelligence security issues. The AI ​​security testing platform launched this time has first fully solved the systematic and automated evaluation of AI algorithms in different scenarios, and provided a security quantitative evaluation indicators for AI industrial -grade applications, which helps to help artificial intelligence algorithms in the industrial world. Safety guarantees are provided in scale applications.

Public information shows that in August this year, this AI security testing platform was a benchmark for the "Trusted Artificial Intelligence Practice Case in 2022" of the Institute. The Ant Group has invested in trusted AI technology research since 2015. It has the "AI Security and Trusted Key Technology" patent with the largest number of globally. The scene is on the ground.

- END -