Recent advances in 2D Human Pose Estimation (HPE) have heightened privacy concerns, an area that remains understudied. Inspired by the disparity in cognitive processing between human vision and neural networks, we address the privacy issue for HPE from the frequency domain perspective for the first time. Specifically, our approach involves applying a discrete cosine transformation to images and then removing some of the lowest and highest frequency channels, followed by training a Convolutional Neural Network (CNN) on randomly selected and randomly ordered frequency components. This method provides a high level of privacy protection by not only obscuring visual content but also reducing image recovery. In the meantime, the HPE model is able to maintain significantly higher performance compared with other privacy preservation approaches in HPE. In addition, we investigate how the selection of frequency components impacts the accuracy and recall of the HPE model. We name this framework PP-Pose. Comparative analyses demonstrate that PP-Pose outperforms existing methods for privacy preservation in HPE and effectively balances privacy protection with pose estimation accuracy.