Improving Stability Estimates in Adversarial Explainable AI through Alternate Search Methods

Explanations of AI models can be manipulated without changing the model's actual decisions. The paper confirms that adversarial attacks can produce drastically different explanations while the underlying model's output remains essentially the same—meaning users could be misled about why an AI made a decision even if the decision itself stays consistent.

Some explanation methods are more resistant to manipulation than others. By measuring how many small changes are needed to destabilize an explanation, researchers can now quantitatively compare which explanation techniques are more reliable, rather than just knowing that vulnerabilities exist.

Stability of explanations matters for trustworthiness. If an explanation can be easily altered through minor tweaks to input data, it undermines user confidence in AI systems. This research provides a way to measure and improve this stability, which is critical for responsible AI deployment in high-stakes applications.

Current "black box" explanation methods have real limitations. Local surrogate methods—tools designed to make complex AI models more understandable—are themselves vulnerable to attack, suggesting that transparency efforts require ongoing scrutiny and improvement rather than being a complete solution to AI interpretability.