pith. sign in

arxiv: 2510.03245 · v2 · submitted 2025-09-25 · 💻 cs.LG · cs.AI· cs.CV

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

Ali Yavari , Alireza Mohamadi , Elham Beydaghi , Philipp Seeb\"ock , Rainer A. Leitgeb This is my paper

Pith reviewed 2026-05-18 14:00 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.CV
keywords explainable AIattribution methodsfrequency domainadversarial perturbationsdeep learningneural networksImageNet evaluation
0
0 comments X

The pith

By selectively perturbing high- and low-frequency image components, FAMPE produces more accurate attribution maps than previous methods.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper proposes a new way to explain deep neural network decisions by focusing on frequency content in images. Instead of perturbing all frequencies equally, FAMPE selectively adjusts high-frequency and low-frequency parts using Fourier transforms and an energy-based cutoff. This selective approach reveals which parts of the spectrum the model depends on most for its predictions. Experiments across different network architectures demonstrate improved performance over existing methods, especially when high-frequency changes are emphasized. The work suggests that careful spectral probing can lead to more accurate and insightful explanations without needing extra baseline choices.

Core claim

The central claim is that by generating adversarial samples that selectively perturb high- and low-frequency components through an FFT-based α-weighted scheme with energy-driven spectral cutoff, and integrating this directly into model parameter exploration, one obtains superior attribution maps for explainability. This translates spectral structure into fine-grained attribution without manual baselines, outperforming prior methods on ImageNet for CNNs and transformers.

What carries the argument

FAMPE's FFT-based α-weighted perturbation scheme, which modulates high- and low-frequency components separately using an energy-driven spectral cutoff to generate attribution signals.

Load-bearing premise

The assumption that frequency-specific perturbations via this scheme produce attribution signals that accurately reflect the model's true decision process without creating misleading artifacts.

What would settle it

If attribution maps generated by FAMPE fail to highlight features that actually influence the model's output when tested with frequency-filtered inputs, that would contradict the central claim.

Figures

Figures reproduced from arXiv: 2510.03245 by Alireza Mohamadi, Ali Yavari, Elham Beydaghi, Philipp Seeb\"ock, Rainer A. Leitgeb.

Figure 1
Figure 1. Figure 1: An illustration of frequency filtering. The top row displays an image separated into its low-frequency (blurred) [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of frequency-aware sample generation. (a) Original image, (b) Original Fourier spectrum, (c) [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of attribution maps from our proposed, FAMPE, and other methods on MaxViT-T and Inception [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Relationship between Cutoff values and α in MaxViT-T [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
read the original abstract

State-of-the-art attribution methods rely on adversarial sample generation that applies an all-pass filter across the frequency spectrum, discarding fine-grained high-frequency information that is demonstrably important for accurate feature attribution in deep neural networks. By generating adversarial samples that selectively perturb high- and low-frequency components, we can probe which spectral features a model relies on most -- directly translating frequency-domain exploration into attribution signals. Building on this insight, we propose FAMPE (Frequency-Aware Model Parameter Explorer), a novel attribution method that introduces an FFT-based \alpha-weighted perturbation scheme -- separately modulating high- and low-frequency components via an energy-driven spectral cutoff -- and, crucially, integrates this frequency-aware exploration directly into model parameter exploration for attribution, a connection that has not been established in prior work. Unlike prior frequency-aware adversarial approaches that target transferability or imperceptibility, FAMPE's specific formulation is designed and validated exclusively for explainability, translating spectral structure into fine-grained attribution maps without requiring any manual baseline selection. Evaluated on ImageNet across four architectures spanning CNNs and Vision Transformers, at fixed \alpha = 0.1 FAMPE outperforms AttEXplore by 4.25% on Inception-v3 and 12.04% on MaxViT-T, with per-sample oracle selection further revealing that low-frequency-dominated images systematically benefit from high-frequency perturbations -- underscoring the potential of adaptive spectral exploration. Our ablation studies confirm that high-frequency perturbations are disproportionately responsible for attribution precision, while excessive low-frequency noise degrades global structural coherence.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes FAMPE, a frequency-aware attribution method for DNNs that generates adversarial samples via an FFT-based α-weighted perturbation scheme. High- and low-frequency components are separately modulated using an energy-driven spectral cutoff, with the resulting signals integrated into model parameter exploration to produce attribution maps. Evaluated on ImageNet for four architectures (CNNs and ViTs), it reports gains over AttEXplore of 4.25% on Inception-v3 and 12.04% on MaxViT-T at fixed α=0.1, with ablations indicating that high-frequency perturbations drive precision while low-frequency noise harms coherence; per-sample oracle selection is used to highlight adaptive benefits.

Significance. If the attribution maps faithfully reflect model reasoning rather than perturbation artifacts, FAMPE could meaningfully advance explainability by moving beyond all-pass filters to explicitly probe spectral dependencies. The explicit link between frequency-domain exploration and attribution is a clear novelty relative to prior frequency-aware adversarial work, and the ablation studies provide useful evidence on the differential roles of high- versus low-frequency components.

major comments (3)
  1. [Abstract] Abstract: the reported 4.25% and 12.04% gains over AttEXplore are stated without error bars, cross-validation details, or statistical tests, so it is impossible to assess whether the improvements are reliable or could arise from variance in the evaluation protocol.
  2. [Abstract] Abstract: the use of 'per-sample oracle selection' to demonstrate benefits of high-frequency perturbations introduces the possibility that reported advantages depend on post-hoc choices rather than the method itself; this directly affects the strength of the central claim that the approach yields valid attribution signals.
  3. [Method] Method formulation (energy-driven spectral cutoff and α-weighted scheme): the cutoff is presented as a fixed component of the perturbation but is not shown to be free of frequency-specific artifacts; because the method deliberately alters the spectral content of the input, the attribution maps could reflect the artificial imbalance rather than the model's original decision process, and no experiment rules this out.
minor comments (2)
  1. [Abstract] The abstract would be clearer if it briefly defined the energy-driven spectral cutoff and stated how α is chosen beyond the fixed value of 0.1.
  2. Notation for the FFT-based perturbation and the precise integration into parameter exploration could be made more explicit to aid reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed comments, which help clarify important aspects of our evaluation protocol and methodological choices. We address each major comment below and indicate the revisions planned for the next version of the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the reported 4.25% and 12.04% gains over AttEXplore are stated without error bars, cross-validation details, or statistical tests, so it is impossible to assess whether the improvements are reliable or could arise from variance in the evaluation protocol.

    Authors: We agree that the abstract reports point estimates without sufficient statistical context. In the revised manuscript we will add error bars derived from multiple independent evaluation runs and include the results of paired statistical significance tests (e.g., Wilcoxon signed-rank test) between FAMPE and AttEXplore. These additions will appear both in the abstract and in an expanded results section. revision: yes

  2. Referee: [Abstract] Abstract: the use of 'per-sample oracle selection' to demonstrate benefits of high-frequency perturbations introduces the possibility that reported advantages depend on post-hoc choices rather than the method itself; this directly affects the strength of the central claim that the approach yields valid attribution signals.

    Authors: The per-sample oracle selection is presented to illustrate the potential value of adaptive spectral exploration, specifically that low-frequency-dominated images benefit from high-frequency perturbations. We acknowledge that this framing risks being read as post-hoc selection that weakens the core claims. In revision we will restrict the abstract to the fixed-α=0.1 results as the primary evidence and move the oracle analysis to a clearly labeled subsection on adaptive strategies, describing it as exploratory rather than central validation of attribution validity. revision: yes

  3. Referee: [Method] Method formulation (energy-driven spectral cutoff and α-weighted scheme): the cutoff is presented as a fixed component of the perturbation but is not shown to be free of frequency-specific artifacts; because the method deliberately alters the spectral content of the input, the attribution maps could reflect the artificial imbalance rather than the model's original decision process, and no experiment rules this out.

    Authors: We recognize the concern that frequency-selective perturbations could introduce artifacts that are not fully disentangled from model reasoning. Our ablation studies already indicate that high-frequency perturbations improve precision while low-frequency components degrade coherence, providing indirect support that the maps capture genuine spectral dependencies. To address the issue more directly we will add a controlled comparison in the revision that contrasts FAMPE attributions with an all-pass perturbation baseline, quantifying differences attributable to the spectral cutoff. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper introduces FAMPE as a novel design that applies an FFT-based α-weighted perturbation with an energy-driven spectral cutoff to generate attribution maps. No load-bearing derivation step is presented that reduces by construction to its own fitted inputs or prior self-citations; the reported gains (4.25% on Inception-v3, 12.04% on MaxViT-T at fixed α=0.1) are empirical comparisons against AttEXplore on ImageNet, and the method's frequency modulation is explicitly framed as a proposed formulation rather than a first-principles prediction derived from the evaluation data itself. The central claims rest on experimental validation rather than any self-referential loop in equations or uniqueness theorems.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The approach rests on the domain assumption that frequency-selective perturbations reveal genuine model reliance and on two tunable elements (α and the spectral cutoff) whose values are chosen rather than derived.

free parameters (2)
  • alpha = 0.1
    Fixed perturbation strength set to 0.1 for the main reported results.
  • energy-driven spectral cutoff
    Threshold separating high- and low-frequency bands, defined via energy but not derived from the target attribution objective.
axioms (1)
  • domain assumption Selective perturbation of frequency components produces attribution signals that accurately reflect the model's internal decision process.
    Invoked when translating spectral exploration into attribution maps.

pith-pipeline@v0.9.0 · 5828 in / 1464 out tokens · 51047 ms · 2026-05-18T14:00:03.396555+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

33 extracted references · 33 canonical work pages · 5 internal anchors

  1. [1]

    Lane, and Cecilia Mascolo

    Petko Georgiev, Sourav Bhattacharya, Nicholas D. Lane, and Cecilia Mascolo. Low-resource multi-task audio sensing for mobile and embedded devices via shared deep neural network representations.Proc. ACM Interact. Mob. Wearable Ubiquitous Technol., 1(3), September 2017

    work page 2017

  2. [2]

    Koppula, Bharad Raghavan, Shane Soh, and Ashutosh Saxena

    Ashesh Jain, Hema S. Koppula, Bharad Raghavan, Shane Soh, and Ashutosh Saxena. Car that knows before you do: Anticipating maneuvers via learning temporal driving models. InProceedings of the 2015 IEEE International Conference on Computer Vision (ICCV), ICCV ’15, page 3182–3190, USA, 2015. IEEE Computer Society

    work page 2015

  3. [3]

    Eunsuk Chong, Chulwoo Han, and Frank C. Park. Deep learning networks for stock market analysis and prediction: Methodology, data representations, and case studies.Expert Systems with Applications, 83:187–205, 2017

    work page 2017

  4. [4]

    T. T. Pham and Y . Shen. A deep causal inference approach to measuring the effects of forming group loans in online non-profit microfinance platform.arXiv preprint arXiv:1706.02795, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  5. [5]

    L. Nie, M. Wang, L. Zhang, S. Yan, B. Zhang, and T. S. Chua. Disease inference from health-related questions via sparse deep learning.IEEE Transactions on Knowledge and Data Engineering, 27(8):2107–2119, 2015

    work page 2015

  6. [6]

    T. Dhar, N. Dey, S. Borra, and R. S. Sherratt. Challenges of deep learning in medical image analysis—improving explainability and trust.IEEE Transactions on Technology and Society, 4:68–75, 2023

    work page 2023

  7. [7]

    S. Ali, T. Abuhmed, S. El-Sappagh, K. Muhammad, J. M. Alonso-Moral, R. Confalonieri, and F. Herrera. Explainable artificial intelligence (xai): What we know and what is left to attain trustworthy artificial intelligence. Information Fusion, 99:101805, 2023

    work page 2023

  8. [8]

    N. Y . Murad, M. H. Hasan, M. H. Azam, N. Yousuf, and J. S. Yalli. Unraveling the black box: A review of explainable deep learning healthcare techniques.IEEE Access, 2024

    work page 2024

  9. [9]

    Maier, J

    T. Maier, J. Menold, and C. McComb. The relationship between performance and trust in ai in e-finance.Frontiers in Artificial Intelligence, 5:891529, 2022

    work page 2022

  10. [10]

    Explainable AI for Mental Disorder Detection via Social Media: A survey and outlook

    Y . Ibrahimov, T. Anwar, and T. Yuan. Explainable ai for mental disorder detection via social media: A survey and outlook.arXiv preprint arXiv:2406.05984, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  11. [11]

    why should i trust you?

    Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. "why should i trust you?": Explaining the predictions of any classifier, 2016

    work page 2016

  12. [12]

    Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra

    Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. InProceedings of the IEEE International Conference on Computer Vision (ICCV), Oct 2017

    work page 2017

  13. [13]

    Score- cam: Score-weighted visual explanations for convolutional neural networks

    Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score- cam: Score-weighted visual explanations for convolutional neural networks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2020

    work page 2020

  14. [14]

    Axiomatic attribution for deep networks, 2017

    Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks, 2017

    work page 2017

  15. [15]

    Explaining deep neural network models with adversarial gradient integration

    Deng Pan, Xin Li, and Dongxiao Zhu. Explaining deep neural network models with adversarial gradient integration. InProceedings of the Thirtieth International Joint Conference on Artificial Intelligence (IJCAI-21), USA, 2021. International Joint Conferences on Artificial Intelligence Organization. Conference Paper

    work page 2021

  16. [16]

    Robust models are more interpretable because attributions look normal.arXiv preprint arXiv:2103.11257, 2021

    Zifan Wang, Matt Fredrikson, and Anupam Datta. Robust models are more interpretable because attributions look normal.arXiv preprint arXiv:2103.11257, 2021

    work page arXiv 2021

  17. [17]

    Mfaba: A more faithful and accelerated boundary-based attribution method for deep neural networks

    Zhiyu Zhu, Huaming Chen, Jiayu Zhang, Xinyi Wang, Zhibo Jin, Minhui Xue, Dongxiao Zhu, and Kim- Kwang Raymond Choo. Mfaba: A more faithful and accelerated boundary-based attribution method for deep neural networks. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 17228–17236, 2024

    work page 2024

  18. [18]

    Zhiyu Zhu, Huaming Chen, Jiayu Zhang, Xinyi Wang, Zhibo Jin, Jason Xue, and Flora D. Salim. AttEXplore: Attribution for explanation with model parameters exploration. InThe Twelfth International Conference on Learning Representations, 2024

    work page 2024

  19. [19]

    Layer-wise Relevance Propagation for Neural Networks with Local Renormalization Layers

    Alexander Binder, Grégoire Montavon, Sebastian Bach, Klaus-Robert Müller, and Wojciech Samek. Layer-wise relevance propagation for neural networks with local renormalization layers.arXiv preprint arXiv:1604.00825, 2016

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  20. [20]

    Learning important features through propagating activation differences

    Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. Learning important features through propagating activation differences. InProceedings of the 34th International Conference on Machine Learning - Volume 70, ICML’17, page 3145–3153. JMLR.org, 2017. 10

    work page 2017

  21. [21]

    Deep inside convolutional networks: Visualising image classification models and saliency maps, 2014

    Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps, 2014

    work page 2014

  22. [22]

    Fast axiomatic attribution for neural networks

    Robin Hesse, Simone Schaub-Meyer, and Stefan Roth. Fast axiomatic attribution for neural networks. InAdvances in Neural Information Processing Systems (NeurIPS), volume 34, pages 19513–19524, 2021

    work page 2021

  23. [23]

    Guided integrated gradients: An adaptive path method for removing noise

    Andrei Kapishnikov, Subhashini Venugopalan, Besim Avci, Ben Wedin, Michael Terry, and Tolga Bolukbasi. Guided integrated gradients: An adaptive path method for removing noise. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5050–5058, 2021

    work page 2021

  24. [24]

    A fourier perspective on model robustness in computer vision

    Dong Yin, Raphael Gontijo Lopes, Jon Shlens, Ekin Dogus Cubuk, and Justin Gilmer. A fourier perspective on model robustness in computer vision. InAdvances in Neural Information Processing Systems (NeurIPS), volume 32, 2019

    work page 2019

  25. [25]

    Haohan Wang, Xindi Wu, Zeyi Huang, and Eric P. Xing. High-frequency component helps explain the general- ization of convolutional neural networks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8684–8694, 2020

    work page 2020

  26. [26]

    Low Frequency Adversarial Perturbation

    Chuan Guo, Jared S. Frank, and Kilian Q. Weinberger. Low frequency adversarial perturbation.arXiv preprint arXiv:1809.08758, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  27. [27]

    J. W. Cooley and J. W. Tukey. An algorithm for the machine calculation of complex fourier series.Mathematics of Computation, 19(90):297–301, 1965

    work page 1965

  28. [28]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In2009 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255. IEEE, 2009

    work page 2009

  29. [29]

    Rethinking the inception architecture for computer vision

    Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2818–2826, 2016

    work page 2016

  30. [30]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770–778, 2016

    work page 2016

  31. [31]

    Very Deep Convolutional Networks for Large-Scale Image Recognition

    Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  32. [32]

    Maxvit: Multi-axis vision transformer

    Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. InEuropean Conference on Computer Vision (ECCV), pages 459–479. Springer, 2022

    work page 2022

  33. [33]

    Rise: Randomized input sampling for explanation of black-box models

    Vitali Petsiuk, Abir Das, and Kate Saenko. Rise: Randomized input sampling for explanation of black-box models. InProceedings of the British Machine Vision Conference (BMVC), 2018. 11

    work page 2018