The Challenge of XAI Trustworthiness

Explainable AI (XAI) methods often suffer from a lack of rigorous validation, leading to explanations that may be misleading or inconsistent. The authors argue that for XAI to be used in high-stakes environments, it requires a formal framework to certify trust. The core problem identified is the gap between an explanation's apparent utility and its actual fidelity to the model's internal logic, as well as its robustness against adversarial input perturbations.

A Formal Audit Methodology

The proposed framework establishes a systematic approach to auditing XAI systems by focusing on two primary metrics: fidelity and robustness.

• Fidelity: This measures how accurately an explanation reflects the model's decision-making process. The framework uses formal verification techniques to ensure that the explanation is not just a post-hoc approximation but a true representation of the model's behavior.
• Robustness: This assesses the stability of explanations when inputs are subjected to minor perturbations. A robust XAI system should produce consistent explanations for similar inputs. The authors define a formal audit process that tests these boundaries, ensuring that small changes in input do not lead to drastic, ungrounded changes in the provided explanation.

From Audit to Certification

The framework moves beyond simple evaluation by providing a path toward 'Trust Certification.' By applying these formal methods, developers can quantify the reliability of their XAI outputs. This allows for the creation of a certification standard where an XAI system is validated against specific thresholds of fidelity and robustness, providing stakeholders with a measurable guarantee of the explanation's quality. This structured approach is designed to move XAI from experimental research into reliable, production-ready applications.