FDCNet: frequency-aware and dynamic curve network for adversarially robust infrared and visible image fusion

Abstract

Infrared and visible image fusion (IVIF) integrates complementary multi-modal information, yet existing methods typically overlook deliberate adversarial attacks. To enhance model robustness in adversarial environments, we propose a novel adversarial attack resilient network, called Frequency-Aware and Dynamic Curve Network for Adversarially Robust Infrared and Visible Image Fusion (FDCNet). Regarding the optimization strategy, we develop a Spatio-Frequency Dynamic Adversarial Loss (LSFDA) to guide robust training by adaptively balancing structural fidelity and high-frequency feature constraints. In terms of network architecture, utilizing a U-Net backbone, we embed a Frequency Awareness Defense Module (FADM) into multi-scale skip connections to intercept cross-layer noise penetration, filtering high-frequency adversarial perturbations while transferring low-level spatial details. These components synergize to achieve effective suppression of adversarial perturbations. However, while successfully suppressing adversarial noise, the two components often exert unavoidable mutual constraints due to their differing priorities. In view of this, we further design an Extrema-guided Dynamic Tone Curve Module (EDTC) at the back-end as a post-compensation mechanism. By combining baseline robust features and multi-source extreme value distributions as spatial priors, this module effectively alleviates the mutual constraints between the LSFDA and the FADM, restoring the pixel distribution of the fused images to balance defense effectiveness and visual quality. Experimental results demonstrate that our FDCNet effectively mitigates the adverse effects of adversarial perturbations, maintaining high-fidelity fusion results and significantly stabilizing the performance of downstream tasks under adversarial attacks. The code is available at https://github.com/00y1/FDCNet. Proposes LSFDA loss to balance structural fidelity and high-frequency constraints. Embeds FADM into skip connections to filter noise and transfer spatial details. Integrates DWT, ACB, and dual-path attention for deep spatio-frequency fusion. Designs EDTC as a post-compensation mechanism to restore pixel distributions. Consistently stabilizes downstream object detection and segmentation under attacks. Proposes LSFDA loss to balance structural fidelity and high-frequency constraints. Embeds FADM into skip connections to filter noise and transfer spatial details. Integrates DWT, ACB, and dual-path attention for deep spatio-frequency fusion. Designs EDTC as a post-compensation mechanism to restore pixel distributions. Consistently stabilizes downstream object detection and segmentation under attacks.