Deep Reinforcement Learning

Case Study

Problem:

AI defenders for critical government cyber systems lack a deep learning system that can continuously improve defense capabilities. This problem is heightened by a lack of advanced AI attackers that can test a diverse range of cyber defenses.

Solution:

Train a robust AI attacker and AI defender to continuously improve cyber defenses.

    • Created a multi-agent system to pit both “Red” AI attackers and “Blue” AI defenders against one another to continually improve both AI cyber capabilities. 
    • Utilize a decision transformer as an innovative solution due to its ability to learn with sparce rewards and large action spaces enabling an unprecedented amount and type of attack scenarios.
    • The AI attacking model presents unique attacking vectors not traditionally seen by cyber defenses.
    • After seeing these unique attack vectors, the robustness of the AI defense model significantly enhances cyber defenses.

Outcome:

    • Successfully trained a “Red” AI attacker which consistently won games against a weak defender after training the AI attacker against a strong defender. 
    • The performance of the AI attacker showed an average increase of 25% compared to the in-domain attacking model.
    • This proved that a multi-agent approach will train both more effective attackers and defenders.