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

Area of Science:

• Artificial Intelligence
• Cybersecurity
• Game Theory

Background:

• Deep learning models face security risks due to limited robustness against adversarial attacks.
• Current research often focuses on single attack methods, neglecting economic impacts and strategic attacker behavior.

Purpose of the Study:

• To develop an economically-grounded security-game framework for evaluating adversarial attacks on deep learning systems.
• To model attacker adaptation using information theory and assess outcomes via economic utilities.

Main Methods:

• Proposed KL-robust Stackelberg Economic Analysis of Risk (KLEAR) framework.
• Modeled attacker's rational adaptation with a KL information budget.
• Analyzed sequential leader-follower interactions with economic evaluations like payback and break-even analysis.

Main Results:

• Derived a closed-form KL-robust best response with a fusion interpretation.
• Quantified attacker incentives on an economic scale, considering costs and gains.
• Identified how information budget and attack prevalence influence attacker and defender strategies through phase diagrams.

Conclusions:

• KLEAR effectively bridges security-game theory and economic evaluation for systematic assessment of adversarial learning.
• The framework enables informed deployment decisions by quantifying economic impacts of security measures.
• Provides insights into regime transitions driven by attacker behavior and system parameters.