An information-theoretic and economic game framework for adversarial robustness evaluation

Area of Science:

• Artificial Intelligence
• Cybersecurity
• Game Theory

Background:

• Deep learning systems face security risks due to limited robustness against adversarial attacks.
• Existing research often overlooks the economic impact and strategic behavior of attackers.

Purpose of the Study:

• To develop a security-game framework that models attacker behavior as resource allocation under realistic constraints.
• To evaluate the economic impact of adversarial attacks on deep learning systems.

Main Methods:

• Proposed KL-robust Stackelberg Economic Analysis of Risk (KLEAR) framework.
• Modeled attacker adaptation using a KL information budget and economic utilities.
• Analyzed sequential leader-follower interactions with defender commitment and attacker mixed strategies.

Main Results:

• Derived a closed-form KL-robust best response with a fusion interpretation.
• Enabled finite-horizon economic evaluation using payback and break-even analysis.
• Quantified incentives on a common economic scale, accounting for costs and gains/losses.

Conclusions:

• KLEAR bridges security-game theory and economic evaluation for systematic assessment of adversarial learning.
• Phase diagrams reveal how information budget and attack prevalence influence attacker composition and defender deployment.
• The framework enables informed deployment decisions by systematically assessing security risks.