AIM: Adversarial Information Masking for Faithfulness Evaluation of Saliency Maps

2026-05-16

AI Summary

Existing saliency map fidelity evaluation methods are prone to unreliable results due to distribution shifts or residual information introduced by masking operations. This work proposes the Adversarial Information Masking (AIM) framework, which introduces adversarial examples for saliency-guided feature replacement and employs complementary masking orders to compare model performance degradation, thereby jointly assessing both saliency map fidelity and mask operator reliability. Reliability is quantified through random attribution bias and the ranking stability of explanation methods, revealing distinct behaviors between signed and unsigned attributions across multiple modalities—including images, audio, and EEG signals. Experiments demonstrate that AIM substantially mitigates mask-induced bias and more effectively discriminates the fidelity performance of different attribution methods.

Abstract

Post-hoc saliency methods are widely used to interpret deep neural networks, but their faithfulness is difficult to evaluate reliably. Existing evaluations mask features according to saliency-induced feature ordering and measure performance degradation, but this degradation can be confounded by the masking operator: zero masking may create out-of-distribution artifacts, while interpolation-based masking may preserve residual predictive information. We propose Adversarial Information Masking (AIM), a saliency-guided adversarial feature replacement framework for evaluating both saliency-map faithfulness and masking-operator reliability. AIM replaces selected features with values from an adversarial counterpart of the input and compares degradation under complementary masking orders. We assess reliability using random-attribution bias and stability of explanation-method faithfulness rankings. Experiments on image, audio, and EEG tasks suggest that AIM reduces masking-induced bias compared with zero and interpolation-based masking, while revealing modality-dependent differences between signed and unsigned attributions.

Problem

• saliency maps
• faithfulness evaluation
• feature masking
• adversarial masking
• post-hoc interpretation

Innovation

• Adversarial Information Masking
• saliency map faithfulness
• feature masking
• attribution reliability
• post-hoc interpretability