Skeleton-based motion data is central to human-computer interaction in virtual, augmented, and mixed reality, with applications spanning clinical decision support, social platforms, and education. However, skeletal representations encode personally identifiable information (PII) through limb proportions, gait dynamics, and other biometric markers. Recent studies show that users can be re-identified with over 80% accuracy and that gender can be predicted with over 87% accuracy from skeleton sequences alone, establishing that raw skeleton data poses a substantial privacy risk. This dissertation addresses that risk through three progressive studies, each relaxing assumptions about attacker knowledge while improving the balance between privacy protection and downstream utility. The first study introduces a white-box, explanation-guided approach. We use Integrated Gradients to compute joint-level sensitivity scores, then apply targeted differential-privacy noise to the most informative joints. Sensitivity-guided noise achieves over 12$\times$ greater re-identification reduction than uniform noise under the same privacy budget, while preserving substantially higher action recognition accuracy. We also present AnonVis, a VR demonstration that visualizes these trade-offs on anonymized avatars. The limitation is that this approach assumes knowledge of the attacker's model architecture and weights. The second study removes that assumption through attack-agnostic motion retargeting. The Privacy-centric deep Motion Retargeting (PMR) model is a CNN-based framework with two encoder instances that share the same architecture but use separate weights, together with cooperative and adversarial classifiers that disentangle motion dynamics from skeletal identity. By transferring a user's motion onto a structurally different target reference skeleton, PMR anonymizes the data without requiring knowledge of the downstream attacker. Experiments show strong re-identification reduction and high motion fidelity, but action recognition accuracy drops, revealing a gap in the privacy-utility trade-off for CNN-based retargeting. The third study closes that gap with Disentangled Transformer Motion Retargeting (DisentangledTMR). DisentangledTMR employs dual embedding streams for motion and identity, a learned gating mechanism that suppresses identity-carrying features, and physical plausibility constraints that enforce bone-length consistency. DisentangledTMR achieves the lowest re-identification rate among retargeting methods while maintaining the highest action recognition accuracy, the lowest reconstruction error, and 4x better bone-length consistency than baselines. Cross-dataset evaluation provided initial evidence of generalization, with re-identification reduced to near-random-chance levels in the tested settings. Together, these three studies trace a clear progression: from targeted noise that requires full attacker knowledge, to CNN-based retargeting that is attack-agnostic but sacrifices utility, to transformer-based disentanglement that achieves the strongest privacy-utility balance with physical plausibility. This dissertation also contributes an overview of skeleton data privacy risks, evaluation protocols, and practical deployment frameworks for privacy-preserving skeleton-based motion systems.
