Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Frequency-Aware Model Parameter Explorer: A new attribution method for improving explainability

Ali Yavari, Alireza Mohamadi, Elham Beydaghi, Rainer A. Leitgeb

TL;DR

This work tackles explainability of DNNs under real-world perturbations by introducing transferable frequency-aware adversarial attacks and the FAMPE attribution method. By decoupling high- and low-frequency perturbations with an alpha-controlled mix and an energy-based cutoff, FAMPE produces more precise attribution maps than prior methods, notably AttEXplore. Across ImageNet experiments on multiple architectures, FAMPE achieves substantial improvements in Insertion Score and demonstrates the value of incorporating frequency-domain perturbations into explainability. The authors also provide ablations to reveal the contributions of each frequency band and plan to release the code for reproducibility and further research.

Abstract

Ensuring the reliability of deep neural networks (DNNs) in the presence of real world noise and intentional perturbations remains a significant challenge. To address this, attribution methods have been proposed, though their efficacy remains suboptimal and necessitates further refinement. In this paper, we propose a novel category of transferable adversarial attacks, called transferable frequency-aware attacks, enabling frequency-aware exploration via both high-and low-frequency components. Based on this type of attacks, we also propose a novel attribution method, named Frequency-Aware Model Parameter Explorer (FAMPE), which improves the explainability for DNNs. Relative to the current state-of-the-art method AttEXplore, our FAMPE attains an average gain of 13.02% in Insertion Score, thereby outperforming existing approaches. Through detailed ablation studies, we also investigate the role of both high- and low-frequency components in explainability.

Frequency-Aware Model Parameter Explorer: A new attribution method for improving explainability

TL;DR

This work tackles explainability of DNNs under real-world perturbations by introducing transferable frequency-aware adversarial attacks and the FAMPE attribution method. By decoupling high- and low-frequency perturbations with an alpha-controlled mix and an energy-based cutoff, FAMPE produces more precise attribution maps than prior methods, notably AttEXplore. Across ImageNet experiments on multiple architectures, FAMPE achieves substantial improvements in Insertion Score and demonstrates the value of incorporating frequency-domain perturbations into explainability. The authors also provide ablations to reveal the contributions of each frequency band and plan to release the code for reproducibility and further research.

Abstract

Ensuring the reliability of deep neural networks (DNNs) in the presence of real world noise and intentional perturbations remains a significant challenge. To address this, attribution methods have been proposed, though their efficacy remains suboptimal and necessitates further refinement. In this paper, we propose a novel category of transferable adversarial attacks, called transferable frequency-aware attacks, enabling frequency-aware exploration via both high-and low-frequency components. Based on this type of attacks, we also propose a novel attribution method, named Frequency-Aware Model Parameter Explorer (FAMPE), which improves the explainability for DNNs. Relative to the current state-of-the-art method AttEXplore, our FAMPE attains an average gain of 13.02% in Insertion Score, thereby outperforming existing approaches. Through detailed ablation studies, we also investigate the role of both high- and low-frequency components in explainability.

Paper Structure

This paper contains 13 sections, 8 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Traditional Methods: Local Approximation and Gradient-based Attribution Methods
  4. Advanced Methods: Adversarial-Sample-Based Attribution Methods
  5. Preliminaries
  6. Fundamental Principles: Sensitivity and Implementation Invariance
  7. IG and AttEXplore methods
  8. Method
  9. Result
  10. Experimental Setup
  11. Experimental Result
  12. Ablation Studies
  13. Conclusion

Figures (4)

  • Figure 1: An illustration of frequency filtering. The top row displays an image separated into its low-frequency (blurred) and high-frequency (edges) components in the spatial domain. The bottom row shows their corresponding Fourier spectra, demonstrating how low-pass and high-pass masks isolate central and outer frequencies, respectively.
  • Figure 2: Illustration of frequency-aware sample generation. (a) Original image, (b) Original Fourier spectrum, (c) Low-frequency image perturbation, (d) High-frequency image perturbation, (e) Reconstructing the image via IFFT to time domain.
  • Figure 3: Comparison of attribution maps from our proposed, FAMPE, and other methods on MaxViT-T and Inception-v3.
  • Figure 4: Relationship between Cutoff values and $\alpha$ in MaxViT-T