Fooling the Image Dehazing Models by First Order Gradient

TL;DR

A group of attack methods based on first order gradient to verify the robustness of the existing dehazing algorithms are designed, which are predicted dehazed image attack, hazy layer mask attack, haze-free image attack and haze-preserved attack.

Abstract

The research on the single image dehazing task has been widely explored. However, as far as we know, no comprehensive study has been conducted on the robustness of the well-trained dehazing models. Therefore, there is no evidence that the dehazing networks can resist malicious attacks. In this paper, we focus on designing a group of attack methods based on first order gradient to verify the robustness of the existing dehazing algorithms. By analyzing the general purpose of image dehazing task, four attack methods are proposed, which are predicted dehazed image attack, hazy layer mask attack, haze-free image attack and haze-preserved attack. The corresponding experiments are conducted on six datasets with different scales. Further, the defense strategy based on adversarial training is adopted for reducing the negative effects caused by malicious attacks. In summary, this paper defines a new challenging problem for the image dehazing area, which can be called as adversarial attack on dehazing networks (AADN). Code and Supplementary Material are available at https://github.com/Xiaofeng-life/AADN Dehazing.

Figures (11)

• Figure 1: Attack results ($L_{P}^{MSE}$) on semantic segmentation dataset Foggy-City, where mACC and mIoU are two metrics for semantic segmentation. "-B" means before attack, and "-A" is after attack.
• Figure 2: The pipelines of adversarial attack and defense on AADN. The contents of the four dotted boxes are (a) original dehazing process, (b) attack dehazing network, (c) adversarial defense training, and (d) illustration of arrows, respectively.
• Figure 3: The optimization direction of attack and defense.
• Figure 4: Dehazing visual results obtained by $A_{N}$. The $\epsilon$ of top row and bottom row are 0 and 8, respectively. Images are from Foggy-City.
• Figure 5: Attack visual results obtained by $L_{P}^{MSE}$ and $L_{M}^{MSE}$. The $P-\epsilon$ and $M-\epsilon$ denote the results when $\epsilon \in \{0, 2, 4, 6, 8\}$. The algorithms of each row from up to down are 4KDehazing and GridDehazeNet, respectively. Images are from O-HAZE.