Federated Learning (FL) protects data privacy by exchanging gradients instead of local training data. However, gradients still pose a risk of being exploited to reconstruct the original data. Traditional attack methods, such as Deep Leakage from Gradients (DLG), optimize dummy gradients to match real gradients but are ineffective under high gradient compression. Modern methods like HCGLA have improved attack performance in this context but fail to fully utilize information from the compressed gradient positions. In this study, we propose a novel approach that leverages information from compressed gradient positions, integrating it into the loss function to enhance attack effectiveness under high compression scenarios. Experiments on CIFAR-100 and LFW datasets demonstrate that our method outperforms the reduced version of HCGLA (which excludes data initialization techniques and denoising models) and DLG, particularly in scenarios involving highly compressed gradients.