Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Fine-Tuning Adversarially-Robust Transformers for Single-Image Dehazing

Vlad Vasilescu, Ana Neacsu, Daniela Faur

TL;DR

This work addresses the vulnerability of single-image dehazing transformers to adversarial perturbations and introduces three lightweight fine-tuning strategies—Last Layer, Scale-and-Bias, and LINEar Adaptation—to boost robustness while preserving clean dehazing quality. The methods are evaluated with adversarial training and TRADES on RESIDE-Outdoor, with tests on remote-sensing datasets, showing that LINEAD provides the strongest robustness under $\ell_\infty$-attacks and that SB offers a favorable trade-off between robustness and computational cost. The findings demonstrate improved resilience to $\ell_0$ perturbations and generalization to out-of-distribution data, offering practical safeguards for remote-sensing deployments. Code for adversarial fine-tuning and attacks is provided, enabling broader application of the proposed robustness enhancements.

Abstract

Single-image dehazing is an important topic in remote sensing applications, enhancing the quality of acquired images and increasing object detection precision. However, the reliability of such structures has not been sufficiently analyzed, which poses them to the risk of imperceptible perturbations that can significantly hinder their performance. In this work, we show that state-of-the-art image-to-image dehazing transformers are susceptible to adversarial noise, with even 1 pixel change being able to decrease the PSNR by as much as 2.8 dB. Next, we propose two lightweight fine-tuning strategies aimed at increasing the robustness of pre-trained transformers. Our methods results in comparable clean performance, while significantly increasing the protection against adversarial data. We further present their applicability in two remote sensing scenarios, showcasing their robust behavior for out-of-distribution data. The source code for adversarial fine-tuning and attack algorithms can be found at github.com/Vladimirescu/RobustDehazing.

Fine-Tuning Adversarially-Robust Transformers for Single-Image Dehazing

TL;DR

This work addresses the vulnerability of single-image dehazing transformers to adversarial perturbations and introduces three lightweight fine-tuning strategies—Last Layer, Scale-and-Bias, and LINEar Adaptation—to boost robustness while preserving clean dehazing quality. The methods are evaluated with adversarial training and TRADES on RESIDE-Outdoor, with tests on remote-sensing datasets, showing that LINEAD provides the strongest robustness under -attacks and that SB offers a favorable trade-off between robustness and computational cost. The findings demonstrate improved resilience to perturbations and generalization to out-of-distribution data, offering practical safeguards for remote-sensing deployments. Code for adversarial fine-tuning and attacks is provided, enabling broader application of the proposed robustness enhancements.

Abstract

Single-image dehazing is an important topic in remote sensing applications, enhancing the quality of acquired images and increasing object detection precision. However, the reliability of such structures has not been sufficiently analyzed, which poses them to the risk of imperceptible perturbations that can significantly hinder their performance. In this work, we show that state-of-the-art image-to-image dehazing transformers are susceptible to adversarial noise, with even 1 pixel change being able to decrease the PSNR by as much as 2.8 dB. Next, we propose two lightweight fine-tuning strategies aimed at increasing the robustness of pre-trained transformers. Our methods results in comparable clean performance, while significantly increasing the protection against adversarial data. We further present their applicability in two remote sensing scenarios, showcasing their robust behavior for out-of-distribution data. The source code for adversarial fine-tuning and attack algorithms can be found at github.com/Vladimirescu/RobustDehazing.

Paper Structure

This paper contains 11 sections, 5 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Adversarial Attacks & Robustness
  3. Adversarial Attacks
  4. Robust Training
  5. Methodology
  6. Fine-tuning for robustness
  7. Measuring Robustness
  8. Experiments & Results
  9. Setup
  10. Results
  11. Conclusions

Figures (4)

  • Figure 1: Small adversarial noise disrupts i2i dehazing networks.
  • Figure 2: Our two proposed fine-tuning methods.
  • Figure 3: Dehazing results for a clean input sample attacked with $\Vert z \Vert_{\infty} \leq \frac{1}{255}$, using different MB-TaylorFormer-B models.
  • Figure 4: Dehazing results for two samples from HazyDet UAV (first row) and RICE-I (second row) attacked with $\Vert z \Vert_{\infty} \leq \frac{1}{255}$.