Mask-Induced Displacement in Audio XAI via Logit Trajectory Decomposition

Masking choices significantly affect AI explanations: The study found that no "fill" (replacement value) used in audio masking is acoustically neutral—zero fill activates silence patterns and Gaussian noise fill activates broadband noise—meaning the method used to mask audio regions fundamentally changes what the AI model responds to, potentially distorting feature importance estimates.

A substantial portion of AI decisions are unexplained by current methods: Between 41-77% of output changes from masking occur through unexpected pathways (off-axis displacement) rather than through the intended mechanism, suggesting that popular audio explanation methods may be missing or misattributing significant portions of model behavior.

Different explanation methods have different blind spots: The paper demonstrates that various attribution methods are "unevenly exposed" to these hidden displacement effects based on their specific sampling and weighting strategies, meaning researchers and practitioners cannot assume all explanation methods are equally reliable or comparable.

Model behavior varies predictably with methodology choices: The off-axis displacement direction remains stable across different masking levels but is specific to each model-fill combination, suggesting that apparent differences in feature importance between models may partly reflect technical choices rather than genuine differences in how models process audio.

Improvements are possible when patterns are understood: Where the unexplained displacement follows stable, low-dimensional patterns, the authors suggest mitigation strategies are feasible, offering a path toward more trustworthy audio AI explanations.