Microsoft Introduces Open Source Machine Learning Security Framework

Machine learning is not only making a leap in the fields of finance, medicine, and military, but also has a great influence on citizens’ lives. However, despite the growing interest in machine learning, Microsoft has released an enterprise open framework (Adversarial ML Threat Matrix) that can detect, respond to, and correct machine learning attacks, saying that countermeasures against cyber attacks on machine learning systems are insufficient.

Microsoft confirmed that cyber attacks targeting commercial machine learning systems are increasing rapidly since 2016, and market research firm Gartner also responded to cyber attacks using AI, and attacks targeting machine learning such as learning data pollution attacks in 2022. It announced a forecast that it will reach 30% of the total.

The risk of cyber attacks on machine learning is increasing day by day, but on the other hand, the sense of crisis about machine learning does not increase. According to a Microsoft survey, 25 out of 28 companies surveyed said they did not have machine learning system security tools. To overcome this phenomenon, Microsoft has released an open source security framework for machine learning in cooperation with MITRE, a non-profit research institute supported by the US government that is known to number the vulnerability identifier CVE.

This tool is intended for use by enterprises with security personnel. It provides a framework for responding to new threats to machine learning systems and adopts a structure similar to the MITRE ATT&CK framework commonly used among security experts. Microsoft and MITER say they are identifying vulnerabilities and attack sets in production environments and investigating the effectiveness of the tool.

Because attacks on machine learning are an important field of research, Microsoft is collecting opinions from experts from the University of Toronto, Cardiff University, and Carnegie Mellon University, and is collecting feedback on this framework at the time of the release of the tool. Related information can be found here .