Deep neural networks (DNNs) for remote sensing image classification and detection often lack adversarial robustness. This paper proposes a frequency-domain projected gradient descent (PGD) attack leveraging the gradient. The image is transformed from the spatial domain to the frequency domain via Fourier transform, and the gradient spectral energy is used to identify key high-frequency regions for generating perturbations, and the perturbations are added to the original image to generate adversarial samples. Using the SSDD dataset and YOLO11 model, we compare adversarial samples from frequency-domain and traditional spatial-domain PGD attacks. Results show that the frequency-domain attack more effectively degrades detection metrics, such as precision and mAP, while producing adversarial samples with higher visual similarity to the originals, demonstrating enhanced attack strength and stealth.