Mask-Induced Displacement in Audio XAI via Logit Trajectory Decomposition

Nico García-Peguinho; David Kelly; Fabrizio Smeraldi; Anna Xambó Sedó

cs.SDSep 27, 2026

Mask-Induced Displacement in Audio XAI via Logit Trajectory Decomposition

Authors: Nico García-Peguinho, David Kelly, Fabrizio Smeraldi, Anna Xambó Sedó

Organizations: School of Electronic Engineering and Computer Science, Queen Mary University of London · Department of Informatics, King's College London · Department of Informatics, King’s College London

Abstract

Perturbation-based XAI methods for audio classifiers often estimate feature importance by masking spectrogram regions and crediting output changes to the retained signal. Yet they typically assume the fill (the mask replacement) is negligible. We propose logit-space trajectory decomposition to examine this assumption. An on-axis component captures output along a line connecting the fully filled (occluded) spectrogram to the fully retained original; an off-axis component captures perpendicular displacement. We evaluate across three fills, three audio classifiers, and 1,100 AudioSet clips. We demonstrate that no fill is acoustically neutral under full occlusion: Zero fill activates silence and Gaussian Noise fill activates broadband noise. Under partial masking, 41-77% of output displacement is off-axis, with the direction of the residual stable across mask retention fraction and specific to each model-fill combination. Attribution methods are unevenly exposed to off-axis displacement through their sampling and weighting strategies, revealing apparatus-dependence. Where the off-axis residual is stable and low-dimensional, its structure affords mitigation.

Figures & tables

Figure 1: Method overview. Four spectrograms (original and three fills) are passed through classifier $\mathcal{F}$ , producing logit vectors $\mathbf{v}\in\mathbb{R}^{527}$ . Each $\mathbf{v}_{\alpha}$ denotes the logit vector at retention fraction $\alpha$ . The logit-space trajectory is decomposed into a scalar on-axis component $\tau$ (along the attribution axis $\mathbf{d}=\mathbf{v}_{\text{orig}}-\mathbf{v}_{\text{foc}}$ ) and a scalar off-axis magnitude $d_{\perp}$ , whose direction is captured by the residual vector $\mathbf{v}_{\perp}\in\mathbb{R}^{527}$ .

Model	Fill	Top-1 class	$\bar{q}$	$\sigma$	$n$	Top-2 class	$\bar{q}$	$\sigma$	$n$
ast+sa	zero	Silence	$.608$	–	1100	Music	$.269$	–	1100
	mean	Mains hum	$.691$	$.220$	857	Hum	$.598$	$.236$	864
	gaussian noise	White noise	$.466$	$.145$	969	Noise	$.176$	$.092$	350
panns-sa	zero	Silence	$.327$	–	1100	Music	$.186$	–	1100
	mean	Buzzer	$.216$	$.123$	273	Hum	$.173$	$.129$	178
	gaussian noise	White noise	$.342$	$.136$	618	Static	$.234$	$.121$	545

Table 1: Top-2 predicted classes per fill at FOC (

N=1{,}100

clips, 22 classes).

\bar{q}

: mean sigmoid confidence.

\sigma

omitted for zero (signal-independent).

	ast+sa	panns-sa	panns+sa
Frying food	Mains hum † – .83	Hum † – .07	Siren § – .13
Drum kit	Mains hum † – .78	Buzzer † – .06	Sound fx – .18
Owl	Sine wave § – .39	Hum † – .09	Sine wave § – .49
Heartbeat	Sine wave § – .37	Music – .12	Sine wave § – .47
Whispering	Mains hum † – .24	Hum † – .09	Sine wave § – .41
Bagpipes	Mains hum † – .14	Buzzer † – .17	Siren § – .14

Table 2: Modal top-1 class and mean

\bar{q}

per source class under mean at FOC. Colours indicate semantic group: † electrical interference (Mains hum, Hum, Buzzer), § tonal periodic (Sine wave, Siren).

Figure 2: Logit-space trajectory analysis across 95.04M partial masks (22 classes, 50 clips, 9,600 masks/clip). Fill: Z = zero, M = mean, G = Gaussian. Row 1: mask distribution and mean on-axis projection $\tau$ by $\alpha$ . Solid line = full logit vector; dotted = GT class ( $\tau_{\text{gt}}$ ); dashed = ideal ( $\tau{=}\alpha$ ). Row 2: per-fill summary table with $\bar{\sigma}_{\boldsymbol{\tau}}$ , $\bar{\sigma}_{\boldsymbol{d_{\perp}}}$ , $d_{\perp}/d_{\text{total}}$ ; mean off-axis residual $d_{\perp}$ by $\alpha$ .

Model	Fill	PR	$\bar{c}$	$\min c_{b,b^{\prime}}$	$\bar{\rho}$ (%)
ast+sa	Z	19.14	.95	.74	18.5
	M	13.37	.94	.76	23.5
	G	15.13	.87	.25	20.0
panns-sa	Z	8.39	.95	.68	32.1
	M	19.56	.86	.20	15.3
	G	16.03	.90	.52	18.5

Table 3: Off-axis residual structure per model–fill condition. PR: participation ratio (effective dimensionality).

\bar{c}

: mean pairwise eigenvector cosine similarity.

\min c_{b,b^{\prime}}

: worst-case similarity.

\bar{\rho}

: variance fraction explained by leading eigenvector. Fill: Z = zero , M = mean , G = gaussian noise .

Explore similar work

APEX: Audio Prototype EXplanations for Classification Tasks
May 11, 2026Piotr Kawa, Kornel Howil, Piotr Borycki +3Audio Understanding Audio Editing
Department of Artificial Intelligence, Wroclaw University of Science and Technology, Poland · IDEAS Research Institute, Poland · Faculty of Mathematics and Computer Science, Jagiellonian University, Poland +1
The Perceived Fragility of Explanations in Audio Models: Manipulation of Attribution with Unchanged Predictions
Jun 12, 2026Piotr Kitłowski, Dominik Wiącek, Mateusz ModrzejewskiAudio Deepfake Detection Explainable AI Methods
Faculty of Electronics and Information Technology, Warsaw University of Technology, Warsaw, Poland.
Listening with Attention: Entropy-Guided Explainability for Transformer-Based Audio Models
Jun 12, 2026Ravi Ranjan, Utkarsh Grover, Xiaomin Lin +1End-To-End Automatic Speech Recognition Models Neural Audio
Florida International University · Miami, USA · University of South Florida +1

May 11, 2026, cs.SDJ/K move · Enter open · S save

APEX: Audio Prototype EXplanations for Classification Tasks

Piotr Kawa, Kornel Howil, Piotr Borycki, Miłosz Adamczyk+2

Department of Artificial Intelligence, Wroclaw University of Science and Technology, Poland · IDEAS Research Institute, Poland · Faculty of Mathematics and Computer Science, Jagiellonian University, Poland+1

Audio Understanding Audio Editing Spectral Features Explainable AI Methods Image Classification

Open paper AskPDF arXiv

Explainable AI (XAI) has achieved remarkable success in image classification, yet the audio domain lacks equally mature solutions. Current methods apply vision-based attribution techniques to spectrograms, overlooking fundamental differences between visual and acoustic signals. While prototype reasoning is promising, acoustic similarity remains multidimensional. We introduce APEX (Audio Prototype EXplanations), a post-hoc framework for interpreting pre-trained audio classifiers. Crucially, APEX requires no fine-tuning of the original backbone and strictly preserves output invariance. APEX disentangles explanations into four perspectives: Square-based prototypes to localize transient events, Time-based for temporal patterns, Frequency-based highlighting spectral bands, and Time-Frequency-based integrating both. This yields intuitive, example-based explanations that respect acoustic properties, providing greater semantic clarity than standard gradient-based methods.

Abstract

Explore similar work

APEX: Audio Prototype EXplanations for Classification Tasks

The Perceived Fragility of Explanations in Audio Models: Manipulation of Attribution with Unchanged Predictions

Listening with Attention: Entropy-Guided Explainability for Transformer-Based Audio Models