A Unified Framework for Evaluating and Enhancing the Transparency of Explainable AI Methods via Perturbation-Gradient Consensus Attribution

XAI methods need comprehensive evaluation beyond accuracy: The paper identifies five critical dimensions—fidelity, interpretability, robustness, fairness, and completeness—that should be assessed together when deploying AI explanations in high-stakes domains like medical imaging and security screening, rather than evaluating explanations in isolation.

Current XAI methods have significant fairness gaps: The proposed PGCA method achieved fairness scores of 4.95±0.03 (on a scale where higher is better), substantially outperforming baseline methods, suggesting that existing explanation techniques may inadvertently provide different quality explanations across demographic groups—a critical concern for equitable AI deployment.

A unified framework enables fair comparison of explanation methods: By standardizing how to measure fidelity, interpretability, robustness, fairness, and completeness through specific metrics (such as Jensen-Shannon divergence for fairness), organizations can objectively compare different XAI approaches and select the most trustworthy for their specific use case.

Combining multiple explanation techniques improves reliability: PGCA fuses two different explanation approaches (perturbation-based and gradient-based methods) to achieve statistically significant improvements across multiple criteria (p < 10⁻⁷), demonstrating that hybrid explanation methods may be more robust than relying on single techniques.

Domain priorities should shape how explanations are evaluated: The framework uses an "entropy-weighted dynamic scoring scheme" that can be adjusted based on what matters most in each application—for example, prioritizing fairness in hiring AI versus fidelity in medical diagnosis—recognizing that one-size-fits-all evaluation is insufficient for responsible AI.