Microsoft is sponsoring a Machine Learning Security Evasion Competition this year, with partners CUJO AI, VMRay, and MRG Effitas, the company has announced.
The competition, which welcomes both machine learning (ML) practitioners and cybersecurity professionals, will allow researchers to exercise their defender and attacker skills, Microsoft says.
The event builds on a similar competition launched at AI Village in August 2019 at DEFCON 27, where contestants participated in a white-box attack against static malware ML models, with many of them identifying approaches that completely and simultaneously bypassed three ML antimalware models.
This year’s contest, the tech company says, is similarly designed to bring to the spotlight countermeasures to adversarial behavior, but also to raise awareness on the various manners in which malware may evade ML systems.
Researchers who score highest at the event are required to publish their detection or evasion strategies in order to win. This year, two different challenges are proposed to researchers.
Set to take place between June 15 and July 23, 2020, the Defender Challenge is meant for the development of novel defenses for attackers to evade. Submissions will be subjected to real-world tests and will have to detect real malware at moderate false-positive rates.
Between August 6 and September 18, 2020, there will be an Attacker Challenge, which will feature a black-box threat model. API access will be provided to hosted antimalware models, including those successfully developed in the Defender Challenge.
Contestants will be required to evade these models using “hard-label” query results and samples from final submissions will be executed in a sandbox to verify their functionality. The final ranking will take into consideration evasion rates and the total number of required API queries.