Differentially Private 2D Human Pose Estimation

TL;DR

This work addresses the privacy risks in 2D human pose estimation by introducing a comprehensive differential privacy framework for 2D-HPE. It combines DP-SGD with a projection-based subspace constraint and a selective privacy scheme (Feature Differential Privacy), culminating in a hybrid Feature-Projective DP method. Across MPII and COCO-based setups, the proposed approach substantially mitigates the utility loss under DP, achieving up to $82.61\%$ mean PCKh@0.5 at $\epsilon=0.8$ and closing the gap to non-private performance, while automatically protecting both subjects and contextual surroundings. The results demonstrate a practical path toward privacy-preserving pose analysis in sensitive applications, with clear guidance on how projection and feature-level privacy interact to balance privacy guarantees and estimation accuracy.

Abstract

Human pose estimation (HPE) has become essential in numerous applications including healthcare, activity recognition, and human-computer interaction. However, the privacy implications of processing sensitive visual data present significant deployment barriers in critical domains. While traditional anonymization techniques offer limited protection and often compromise data utility for broader motion analysis, Differential Privacy (DP) provides formal privacy guarantees but typically degrades model performance when applied naively. In this work, we present the first comprehensive framework for differentially private 2D human pose estimation (2D-HPE) by applying Differentially Private Stochastic Gradient Descent (DP-SGD) to this task. To effectively balance privacy with performance, we adopt Projected DP-SGD (PDP-SGD), which projects the noisy gradients to a low-dimensional subspace. Next, we incorporate Feature Differential Privacy(FDP) to selectively privatize only sensitive features while retaining public visual cues. Finally, we propose a hybrid feature-projective DP framework that combines both approaches to balance privacy and accuracy for HPE. We evaluate our approach on the MPII dataset across varying privacy budgets, training strategies, and clipping norms. Our combined feature-projective method consistently outperforms vanilla DP-SGD and individual baselines, achieving up to 82.61\% mean PCKh@0.5 at $ε= 0.8$, substantially closing the gap to the non-private performance. This work lays foundation for privacy-preserving human pose estimation in real-world, sensitive applications.

Figures (4)

• Figure 1: Overview of our private HPE pipeline coupling a TinyViT based backbone with Coordinate Classification Keypoint head. The Public feature batch $B_{pub}$ is generated from the private batch $B_{priv}$ using $\psi$ both of which are given as input in a single iteration. Additionally, a public image set $B_{pub}^{proj}$ independent of $B_{priv}$ is used to calculate public gradients for projection at specified intervals. Red Arrow indicates the propogation of private gradients and Green Arrow indicates propagation of public feature gradients. Blue Dotted Arrow indicates the propogation of cumulative denoised gradient from Feature Projective DP.
• Figure 2: Comparison of PCKh@0.5 across private and non-private methodologies under different training strategies with varied privacy budget ($\epsilon$) and clipping thresholds ($C$).
• Figure 3: Depiction of qualitative results on DP-SGD, Projection DP-SGD and Feature Projection DP-SGD. We specifically show results on Finetuning with $C=0.1$ at various privacy budgets.
• Figure 4: Figures (a-e)Depiction of qualitative results on DP-SGD, Projection DP-SGD and Feature Projection DP-SGD. We specifically show results on Finetuning with $C=0.1$ at various privacy budgets. (f) Representation of Raw (Private) image compared to public feature (gaussian blurred).