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

TL;DR

This work addresses the lack of objective faithfulness evaluation for XAI explanations in spectrogram-based machine anomalous sound detection. It introduces a frequency-band perturbation framework that links attribution relevance to actual model sensitivity by correlating band-wise attribution with prediction changes when bands are removed, using Spearman correlation and Fisher z-transforms. Across the DCASE 2023 Task 2 dataset, Occlusion consistently exhibits the strongest alignment with model behaviour (approximately $ar{ ho}_s \,\approx\,0.884$), while Integrated Gradients shows moderate reliability ($\bar{\rho}_s \approx 0.530$) and Grad-CAM/SmoothGrad perform more variably or poorly. The results reveal a robust low-frequency reliance in ASD cues and demonstrate the value of a reproducible, quantitative framework for benchmarking audio explanations, guiding more trustworthy interpretation and design of spectrogram-based ASD systems.

Abstract

Explainable AI (XAI) is commonly applied to anomalous sound detection (ASD) models to identify which time-frequency regions of an audio signal contribute to an anomaly decision. However, most audio explanations rely on qualitative inspection of saliency maps, leaving open the question of whether these attributions accurately reflect the spectral cues the model uses. In this work, we introduce a new quantitative framework for evaluating XAI faithfulness in machine-sound analysis by directly linking attribution relevance to model behaviour through systematic frequency-band removal. This approach provides an objective measure of whether an XAI method for machine ASD correctly identifies frequency regions that influence an ASD model's predictions. By using four widely adopted methods, namely Integrated Gradients, Occlusion, Grad-CAM and SmoothGrad, we show that XAI techniques differ in reliability, with Occlusion demonstrating the strongest alignment with true model sensitivity and gradient-+based methods often failing to accurately capture spectral dependencies. The proposed framework offers a reproducible way to benchmark audio explanations and enables more trustworthy interpretation of spectrogram-based ASD systems.

Figures (5)

• Figure 1: Framework for evaluating methods on a machine audio anomaly detection task.
• Figure 2: Spectrogram representation of an audio signal, with brighter colours representing higher magnitude sounds and darker colours representing lower magnitude sounds.
• Figure 3: Visualisation of methods on audio spectrograms. Each column represents a different sample from the test set. The first row shows the original audio spectrograms, and the subsequent rows represent the relevance attributions of the methods in the following order: Integrated Gradients (second row), Occlusion (third row), Grad-CAM (fourth row), and SmoothGrad (fifth row). Regions of interest (highlighted in purple) were generated using the methods described in this study.
• Figure 4: AUC results for each machine with different frequency bands on the development subset
• Figure 5: AUC results for each machine with different frequency bands on the evaluation subset