Shortcut Learning in Binary Classifier Black Boxes: Applications to Voice Anti-Spoofing and Biometrics

TL;DR

The paper tackles the problem of dataset bias and shortcut learning in binary classifiers by introducing a unified framework that analyzes how training and test data influence classifier scores beyond traditional error rates. It builds a two-pronged methodology—interventional perturbations and observational nuisance analysis—connected by linear mixed-effects modeling to quantify how data-related factors shift scores in black-box detectors. The authors validate the approach on two speech tasks: anti-spoofing and speaker verification, showing that shortcuts can act as discriminative cues under certain configurations and highlighting the need for bias-aware evaluation. The work contributes a transparent, model-agnostic post-hoc analysis tool for bias discovery, with broad implications for explainable AI and fair, reliable data-driven decision-making across domains.

Abstract

The widespread adoption of deep-learning models in data-driven applications has drawn attention to the potential risks associated with biased datasets and models. Neglected or hidden biases within datasets and models can lead to unexpected results. This study addresses the challenges of dataset bias and explores ``shortcut learning'' or ``Clever Hans effect'' in binary classifiers. We propose a novel framework for analyzing the black-box classifiers and for examining the impact of both training and test data on classifier scores. Our framework incorporates intervention and observational perspectives, employing a linear mixed-effects model for post-hoc analysis. By evaluating classifier performance beyond error rates, we aim to provide insights into biased datasets and offer a comprehensive understanding of their influence on classifier behavior. The effectiveness of our approach is demonstrated through experiments on audio anti-spoofing and speaker verification tasks using both statistical models and deep neural networks. The insights gained from this study have broader implications for tackling biases in other domains and advancing the field of explainable artificial intelligence.

Figures (2)

• Figure 1: We model shortcut learning effect in binary classification tasks through two complementary approaches. The interventional method perturbs (e.g. adds noise to) an existing dataset to introduce systematic covariate shift to the class-conditional data distributions either on the training or test side. In the observational approach we assume the presence of systematic covariate shifts and extract relevant nuisance feature(s) $\ell_i$ (such as signal-to-noise ratio) that attempt to predict the class label based on the irrelevant features. In both approaches we model the dependency of the classifier output score $s_i$ (on either the intevention parameters or the nuisance score) using linear mixed effects (LME) modeling.
• Figure 2: Heatmaps of bonafide-spoof EER for five different interventions and two spoofing countermeasures (LFCC-GMM sahidullah15_interspeech, AASIST with RawBoost tak2022rawboost). The three sets of panels correspond to different training data interventions (a: no interventions; b: intervention to bona fide class only; c: intervention to spoof training only), and each of the fifteen panels displays the results for varied degrees of test-side intervention probability.