Reproduce “Fourier-Based Augmentations for Improved Robustness and Uncertainty Calibration” on OpenResearch

About this paper

Ryan Soklaski, Michael Yee, Theodoros Tsiligkaridis

Researchers from MIT Lincoln Laboratory developed AugSVF, an extension of the AugMix data augmentation framework, by integrating Fourier-basis perturbations to enhance model robustness and uncertainty calibration. The approach achieves superior performance on common corruption benchmarks and dramatically reduces susceptibility to Fourier-based adversarial attacks, without incurring typical trade-offs in overall robustness.

Key insights

• The AugMix framework can be effectively tailored with specific augmentation primitives, such as Fourier-basis perturbations, to target particular vulnerabilities without compromising general robustness.
• Improving robustness against fundamental Fourier-basis attacks simultaneously enhances overall model accuracy and uncertainty calibration on diverse common corruptions.
• The study validates that a Fourier perspective offers a powerful approach for understanding and mitigating model vulnerabilities, guiding the design of more robust augmentation strategies.

Summary by alphaXiv.