MUFIA

Frequency-Based Vulnerability Analysis of Deep Learning Models against Image Corruptions

MUFIA (Multiplicative Frequency Attack) is a framework for analyzing the vulnerability of deep learning models to image corruptions through the lens of frequency analysis. The project provides tools to:

  1. Analyze how different frequency components in images contribute to model predictions
  2. Evaluate model robustness against various types of image corruptions
  3. Understand the relationship between frequency characteristics and model vulnerability

Key Features

  • Frequency Analysis Tools: Decompose images into different frequency components using DCT transforms
  • Corruption Analysis: Evaluate model performance against common image corruptions
  • Visualization Tools: Generate visualizations to understand frequency-based vulnerabilities
  • Model Testing Framework: Test different CNN architectures for robustness analysis
Overview of the MUFIA framework for analyzing model vulnerabilities through frequency analysis.

You can find more details in our paper (Machiraju et al., 2023) and check out the code on GitHub.

References

2023

  1. letterM.png
    Frequency-Based Vulnerability Analysis of Deep Learning Models against Image Corruptions
    Harshitha Machiraju , Michael H Herzog , and Pascal Frossard
    (Under Submission), 2023
    arXiv