Cyber Autonomy Gym for Experimentation (CAGE) – DST Group calls for submissions
DST Group has issued a Challenge to researchers in Autonomous Cyber Operations (ACO) which closes on 1 February, 2022.
Recent advances in Artificial Intelligence (AI) technologies show promise for ACO, offering the potential for distributed, adaptive defensive measures at machine speed and scale. The cyber domain is a particularly challenging domain for autonomous AI.
DST Group has nominated a challenge in this space that it believes requires further research in order to enable ACO to become an operational capability. To facilitate this AI research, the TTCP CAGE working group are releasing CybORG, an experimental platform using the OpenAI Gym interface together with a cyber security scenario and a challenge to which it invites researchers to respond.
DST Group says, “Our aim is to support the development of AI tactics, techniques and procedures with CybORG and a series of CAGE scenarios with associated challenge problems in order to support practical demonstrations of ACO. We wish to engage the AI and cyber security research communities, especially to leverage domain experts outside of the cyber field, and by encapsulating the cyber elements in environments such as CybORG along with the CAGE scenarios and challenge problems, we hope that the cyber problem set becomes accessible to a wider audience. The first CAGE scenario and associated challenge problem were released at the IJCAI-21 1st International Workshop on Adaptive Cyber Defense (ACD 2021).”
The CAGE challenge environment, CybORG, and first challenge are available here. Any enquiries can be directed to CAGEACOChallenge@dst.defence.gov.au
The CAGE challenge is written in Python. Dependencies can be installed using pip. Further instructions are included on the GitHub page. The challenge includes red agents to test against and an example blue agent. Submissions should implement the same methods of the example blue agent.
Important dates
20 August 2021: Challenge announced at ACD 2021 and open for submissions.
November 2021: Release of extended version of the Challenge problem with a Misinform action available to the blue agent.
1 February 2022: Challenge closed.
4 February 2022: Final results announced.
28 February 2022: Final results presented at AAAI-22 Workshop on Artificial Intelligence for Cyber Security (AICS).