ABLE: Using Adversarial Pairs to Construct Local Models for Explaining Model Predictions

TL;DR

<3-5 sentence high-level summary> ABLE tackles the instability and fidelity gaps of local explanation methods by constructing adversarial pairs that bracket the local decision boundary around a test instance. It generates neighborhood points, creates minimal perturbations on both sides of the boundary, and trains a simple surrogate to faithfully approximate the boundary neighborhood. Across six UCI datasets and three target architectures, ABLE demonstrates superior Fidelity and Stability compared to LIME, GLIME, USLIME, and CALIME, with ABLE_PGD often providing the best performance. The work also provides a public implementation and outlines directions for extending the approach to other data modalities and perceptual-aware perturbations.

Abstract

Machine learning models are increasingly used in critical applications but are mostly "black boxes" due to their lack of transparency. Local explanation approaches, such as LIME, address this issue by approximating the behavior of complex models near a test instance using simple, interpretable models. However, these approaches often suffer from instability and poor local fidelity. In this paper, we propose a novel approach called Adversarially Bracketed Local Explanation (ABLE) to address these limitations. Our approach first generates a set of neighborhood points near the test instance, x_test, by adding bounded Gaussian noise. For each neighborhood point D, we apply an adversarial attack to generate an adversarial point A with minimal perturbation that results in a different label than D. A second adversarial attack is then performed on A to generate a point A' that has the same label as D (and thus different than A). The points A and A' form an adversarial pair that brackets the local decision boundary for x_test. We then train a linear model on these adversarial pairs to approximate the local decision boundary. Experimental results on six UCI benchmark datasets across three deep neural network architectures demonstrate that our approach achieves higher stability and fidelity than the state-of-the-art.

Figures (2)

• Figure 1: In this example, $x_{\text{test}}$ is located far away from the global decision boundary. The local model approximates the behavior of the target model in a small neighborhood of $x_{\text{test}}$, in terms that it produces the same classifications for the neighborhood data points as the target model. However, the local model (purple line) does not align with the local decision boundary (red curve) of the target model for $x_{\text{test}}$.
• Figure 2: (I) $X_{test}$ is surrounded by neighborhood points (blue dots for same class and green boxes for different class) within radius $r$ (red: global decision boundary). (II) $X_{test'}$ is the projection onto the decision boundary; forward and reverse adversarial attacks generate adversarial pairs. (III) Adversarial pairs (blue crosses for the same class and green crosses for a different class) bracket the local decision boundary and are used to train the local model.