Skeleton-based representations are integral to modeling human motion in computer vision due to their efficiency and effectiveness. However, despite the abstraction provided by skeletal data, privacy concerns persist as sensitive personal information can still be inferred from motion patterns. In this paper, we propose a novel methodology that leverages explainable AI techniques to enhance privacy preservation in skeleton data. By computing joint importance scores using Integrated Gradients, we identify sensitive joints and apply targeted privacy-preserving transformations, such as masking and differential privacy mechanisms, to obfuscate private information while minimizing the impact on action recognition performance. Our approach effectively reduces re-identification risks while preserving the utility of the data, as demonstrated through extensive evaluations on benchmark datasets. The code for our implementation will be publicly available alongside the publication of this paper.