A Framework for Evaluating Faithfulness in Explainable AI for Machine Anomalous Sound Detection Using Frequency-Band Perturbation

Not all AI explanations are equally trustworthy: The paper tested four popular explanation methods and found significant differences in their reliability. Occlusion performed best, while gradient-based methods (Integrated Gradients, Grad-CAM, SmoothGrad) often failed to accurately identify which sound frequencies the AI model actually used to make decisions.

Visual explanations alone aren't enough: Current practice relies on researchers visually inspecting explanation maps, but this subjective approach cannot verify whether the highlighted regions truly reflect what the AI "learned." The paper demonstrates the need for objective testing methods to validate explanations before trusting them.

A new testing framework improves accountability: By systematically removing different frequency bands from audio and measuring how the AI's predictions change, researchers can now objectively measure whether an explanation method is faithful. This creates a reproducible benchmark for evaluating explanation quality in sound-detection systems.

Implications for real-world deployment: In applications like industrial equipment monitoring or safety systems that rely on anomalous sound detection, using unreliable explanations could lead to misplaced trust in AI decisions. This framework helps ensure that when operators rely on AI explanations to understand why an alert was triggered, those explanations actually reflect the model's true reasoning.