This work explores the impact of defender assumptions about attacker knowledge on the performance of automated cybersecurity defense agents. The key findings are:
Defenders who assume the attacker has complete knowledge of the system perform worse than defenders who assume the attacker has limited knowledge. This "price of pessimism" leads to suboptimal policy convergence for the defending agent.
Defending agents trained against learning attackers are highly capable against algorithmic attackers, even if they have not seen those algorithmic attackers during training.
The authors introduce a novel use of the Bayes-Hurwicz criterion to parameterize attacker decision-making under uncertainty, and demonstrate its impact on attacker and defender performance.
The authors extend the YAWNING-TITAN reinforcement learning framework to enable independent training of attacking and defending agents.
The results highlight the importance of accurately modeling attacker knowledge when developing automated cybersecurity defenses. Overestimating the attacker's capabilities can lead to suboptimal defensive strategies, while underestimating the attacker can leave the system vulnerable. The authors recommend that future work in this area should carefully consider the likely knowledge and learning dynamics of real-world attackers.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Erick Galink... at arxiv.org 10-01-2024
https://arxiv.org/pdf/2409.19237.pdfDeeper Inquiries