What we know
In the simulated scenarios, more decoys reduced the attacker agent’s episode win percentage and increased the steps needed to win.
Evidence limits
The setting is a toy network with a predefined reward function. No autonomous defender, uncertainty intervals or field outcomes are reported; the curves do not establish effectiveness in real networks.
Reviewed: Full text.
Full-text review · 2026-10-02
Design and scope
CyberBattleSim ToyCTF simulation with two attacker types and DQN and tabular Q-learning agents; compares decoy counts with an explicit penalty for connecting to a decoy.
Main finding
In the simulated scenarios, more decoys reduced the attacker agent’s episode win percentage and increased the steps needed to win.
Evidence limits
The setting is a toy network with a predefined reward function. No autonomous defender, uncertainty intervals or field outcomes are reported; the curves do not establish effectiveness in real networks.
Locators: PDF pp. 7–10, Sections 6.1–7, Figures 2–3
The full text was examined and locators were retained. Single-reviewer synthesis without independent extraction or formal quality appraisal.
Sources and provenance
- Cyber Deception Metrics For Interconnected Complex Systems
Full text · 2026-10-02 - Full text examined for Atlas review
Full text · 2026-10-02
Reviewed: 2026-10-02. This record may change when new evidence is found.