Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Video-DPRP: A Differentially Private Approach for Visual Privacy-Preserving Video Human Activity Recognition

Allassan Tchangmena A Nken, Susan Mckeever, Peter Corcoran, Ihsan Ullah

TL;DR

This work tackles privacy concerns in video-based HAR by proposing Video-DPRP, a per-video sample-wise differentially private random-projection framework that enables both DP guarantees and direct visual privacy assessment on video data. The method reshapes each video to a 2D representation, applies a Johnson-Lindenstrauss random projection, and injects DP noise into the projected data and its covariance; the noisy covariance’s right singular vectors guide a reconstruction that yields a visually private video. The approach is evaluated on HAR datasets (UCF101, HMDB51) and visual privacy datasets (PA-HMDB, VISPR), and is compared against DP-SGD and SOTA visual privacy methods. Results show Video-DPRP can maintain competitive HAR performance while providing robust visual privacy, with a controllable privacy-utility trade-off governed by $(\epsilon,\delta)$ and the budget allocator $b$, offering a model-free alternative that bridges DP theory and practical privacy in video HAR.

Abstract

Considerable effort has been made in privacy-preserving video human activity recognition (HAR). Two primary approaches to ensure privacy preservation in Video HAR are differential privacy (DP) and visual privacy. Techniques enforcing DP during training provide strong theoretical privacy guarantees but offer limited capabilities for visual privacy assessment. Conversely methods, such as low-resolution transformations, data obfuscation and adversarial networks, emphasize visual privacy but lack clear theoretical privacy assurances. In this work, we focus on two main objectives: (1) leveraging DP properties to develop a model-free approach for visual privacy in videos and (2) evaluating our proposed technique using both differential privacy and visual privacy assessments on HAR tasks. To achieve goal (1), we introduce Video-DPRP: a Video-sample-wise Differentially Private Random Projection framework for privacy-preserved video reconstruction for HAR. By using random projections, noise matrices and right singular vectors derived from the singular value decomposition of videos, Video-DPRP reconstructs DP videos using privacy parameters ($ε,δ$) while enabling visual privacy assessment. For goal (2), using UCF101 and HMDB51 datasets, we compare Video-DPRP's performance on activity recognition with traditional DP methods, and state-of-the-art (SOTA) visual privacy-preserving techniques. Additionally, we assess its effectiveness in preserving privacy-related attributes such as facial features, gender, and skin color, using the PA-HMDB and VISPR datasets. Video-DPRP combines privacy-preservation from both a DP and visual privacy perspective unlike SOTA methods that typically address only one of these aspects.

Video-DPRP: A Differentially Private Approach for Visual Privacy-Preserving Video Human Activity Recognition

TL;DR

This work tackles privacy concerns in video-based HAR by proposing Video-DPRP, a per-video sample-wise differentially private random-projection framework that enables both DP guarantees and direct visual privacy assessment on video data. The method reshapes each video to a 2D representation, applies a Johnson-Lindenstrauss random projection, and injects DP noise into the projected data and its covariance; the noisy covariance’s right singular vectors guide a reconstruction that yields a visually private video. The approach is evaluated on HAR datasets (UCF101, HMDB51) and visual privacy datasets (PA-HMDB, VISPR), and is compared against DP-SGD and SOTA visual privacy methods. Results show Video-DPRP can maintain competitive HAR performance while providing robust visual privacy, with a controllable privacy-utility trade-off governed by and the budget allocator , offering a model-free alternative that bridges DP theory and practical privacy in video HAR.

Abstract

Considerable effort has been made in privacy-preserving video human activity recognition (HAR). Two primary approaches to ensure privacy preservation in Video HAR are differential privacy (DP) and visual privacy. Techniques enforcing DP during training provide strong theoretical privacy guarantees but offer limited capabilities for visual privacy assessment. Conversely methods, such as low-resolution transformations, data obfuscation and adversarial networks, emphasize visual privacy but lack clear theoretical privacy assurances. In this work, we focus on two main objectives: (1) leveraging DP properties to develop a model-free approach for visual privacy in videos and (2) evaluating our proposed technique using both differential privacy and visual privacy assessments on HAR tasks. To achieve goal (1), we introduce Video-DPRP: a Video-sample-wise Differentially Private Random Projection framework for privacy-preserved video reconstruction for HAR. By using random projections, noise matrices and right singular vectors derived from the singular value decomposition of videos, Video-DPRP reconstructs DP videos using privacy parameters () while enabling visual privacy assessment. For goal (2), using UCF101 and HMDB51 datasets, we compare Video-DPRP's performance on activity recognition with traditional DP methods, and state-of-the-art (SOTA) visual privacy-preserving techniques. Additionally, we assess its effectiveness in preserving privacy-related attributes such as facial features, gender, and skin color, using the PA-HMDB and VISPR datasets. Video-DPRP combines privacy-preservation from both a DP and visual privacy perspective unlike SOTA methods that typically address only one of these aspects.

Paper Structure

This paper contains 14 sections, 3 theorems, 3 equations, 2 figures, 5 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method Overview
  4. Video Transformation
  5. Differential Privacy
  6. Random Projection
  7. Video-DPRP Algorithm
  8. Privacy Guarantee of Video-DPRP
  9. Experiments
  10. Datasets
  11. Implementation details
  12. Evaluation Metrics and Protocols
  13. Ablation Study
  14. Conclusion

Key Result

Lemma 2

Let $\mathcal{S}$ be a set of $n$ points such that $\mathcal{S} \subset \mathbb{R}^d$, with $\lambda > 0$ and $k = \frac{20 \log n}{\lambda^2}$. There exists a Lipschitz mapping $f: \mathbb{R}^d \to \mathbb{R}^k$ that distorts all pairwise distances by a factor of $1 \pm \lambda$. For any $x, y \in

Figures (2)

  • Figure 1: In $(a)$, privacy is ensured during training (in-training) using differential privacy (DP), but not directly on the video itself. As a result visual privacy cannot be assessed. In $(b)$, the video is transformed prior to training using either obfuscation methods or adversarial approaches, but the privacy-utility trade-off cannot be quantify as clearly as in DP. In $(c)$ (ours), privacy is ensured using DP, directly on the video. This approach allows for visual privacy evaluation, where privacy-utility trade-off is quantified using the $\epsilon$,$\delta$ parameters of DP.
  • Figure 2: Video-DPRP consists of the following components:$(1)$ Each video frame is reshaped and flattened, then concatenated to form a video $X$ of dimension $(T,w\times h\times 3)$. $(2)$ A random projection matrix $\mathcal{R}_{\mathcal{N}(0,\sigma_{p})}$ reduces $X$ to a lower-dimensional space $(T,k)$. $(3)$ Noise is added to both the projected video and its covariance matrix, from which the right singular component $V$ of the noisy covariance $Q$ is used to reconstruct a differentially private video (see Section \ref{['overview']} for details).

Theorems & Definitions (4)

  • Definition 1: Differential Privacy
  • Lemma 2: Johnson-Lindenstrauss johnson1984extensions
  • Theorem 3: Privacy of projected video $P$
  • Theorem 4: Privacy of covariance matrix $P_{cov}$