Reproduce “AIM: Adversarial Information Masking for Faithfulness Evaluation of Saliency Maps” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Chia-Ying Hsieh, Hsin-Yuan Fang, Chun-Shu Wei

This work introduces Adversarial Information Masking (AIM), a new framework for evaluating saliency map faithfulness that reduces masking-induced bias in deep neural network explanations. AIM replaces masked features with values from an adversarial counterpart, demonstrating improved reliability and consistency across image, audio, and EEG tasks compared to conventional masking methods.

Key insights

  • Masking-induced bias is a critical confounder in saliency faithfulness evaluation, significantly influencing conclusions about which explanation method is "most faithful."
  • Adversarial feature replacement reduces this bias, leading to more reliable and consistent assessments of saliency map quality across diverse data modalities.
  • The interpretation of attribution sign versus magnitude is modality-dependent; unsigned magnitude may better capture model sensitivity in temporal signals, while signed values can complicate spectral interpretation.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.