Rigorous Feature Importance Scores based on Shapley Value and Banzhaf Index

Feature importance scores can now account for adversarial robustness: Traditional methods for explaining which features matter most in AI decisions have overlooked non-explanatory feature sets, but this research shows that considering these sets helps identify which features are most effective at preventing adversarial examples—inputs deliberately designed to fool AI systems.

Game theory provides a rigorous foundation for AI explanations: By applying Shapley value and Banzhaf index (established game theory concepts), this work offers mathematically rigorous approaches to feature attribution that go beyond previous logic-based explanation methods, potentially making AI explanations more trustworthy for high-stakes applications like healthcare or criminal justice.

The relationship between explanations and adversarial vulnerabilities matters: The paper reveals an important connection between formal explanations and adversarial examples, suggesting that understanding why an AI makes a decision is linked to understanding how robust that decision is to manipulation—a critical consideration for responsible AI deployment.

Computational complexity is a practical consideration: While the paper identifies important properties of the proposed scores, it also studies their computational complexity, acknowledging that practitioners need to understand the practical trade-offs between explanation rigor and computational feasibility when implementing these methods in real systems.