Only Whats Necessary: Pareto Optimal Data Minimization for Privacy Preserving Video Anomaly Detection

Nazia Aslam; Abhisek Ray; Thomas B. Moeslund; Kamal Nasrollahi

cs.CVMar 27, 2026

Only Whats Necessary: Pareto Optimal Data Minimization for Privacy Preserving Video Anomaly Detection

Authors: Nazia Aslam, Abhisek Ray, Thomas B. Moeslund, Kamal Nasrollahi

Organizations: Aalborg University, Denmark · Aarhus University, Denmark · Milestone System, Denmark

Abstract

Video anomaly detection (VAD) systems are increasingly deployed in safety critical environments and require a large amount of data for accurate detection. However, such data may contain personally identifiable information (PII), including facial cues and sensitive demographic attributes, creating compliance challenges under the EU General Data Protection Regulation (GDPR). In particular, GDPR requires that personal data be limited to what is strictly necessary for a specified processing purpose. To address this, we introduce Only What's Necessary, a privacy-by-design framework for VAD that explicitly controls the amount and type of visual information exposed to the detection pipeline. The framework combines breadth based and depth based data minimization mechanisms to suppress PII while preserving cues relevant to anomaly detection. We evaluate a range of minimization configurations by feeding the minimized videos to both a VAD model and a privacy inference model. We employ two ranking based methods, along with Pareto analysis, to characterize the resulting trade off between privacy and utility. From the non-dominated frontier, we identify sweet spot operating points that minimize personal data exposure with limited degradation in detection performance. Extensive experiments on publicly available datasets demonstrate the effectiveness of the proposed framework.

Figures & tables

Panel of Figure 1 — Figure 1: Overview of the proposed data minimization framework. Given an input video clip $x\in\mathbb{R}^{T\times H\times W\times C}$ , a configurable minimization module $T_{\theta}$ produces a reduced representation $z=T_{\theta}(x)$ . The minimized video is then evaluated along two downstream objectives: anomaly detection using a VAD model $f_{\phi}$ and privacy leakage using a privacy inference model $g_{\psi}$ .

\mathcal{P}=\left\{S_{i}\in\mathcal{S}\;\middle|\;\nexists\,s_{j}\in\mathcal{S}\text{ such that }A_{j}\geq A_{i},F_{j}\leq F_{i},C_{j}\leq C_{i}\right\}

Algorithm 1 Selection of Pareto-optimal settings

Method	Setting	UCSD Ped1	UCSD Ped2	CUHK Avenue	PAHMDB
Method	Setting	AUC ( $\uparrow$ %)	AUC ( $\uparrow$ %)	AUC ( $\uparrow$ %)	cMAP ( $\downarrow$ %)	F1 ( $\downarrow$ %)
Raw	–	79.10	92.49	83.01	70.20	0.396
Temporal sampling	$s=5$	72.98	85.87	77.06	68.98	0.329
Temporal sampling	$s=10$	67.49	84.79	73.43	68.24	0.321
Downsample $\times 2$	–	68.98	67.07	73.86	65.34	0.321
Downsample $\times 4$	–	65.21	63.40	72.17	64.32	0.301

Table 1: Performance under breadth- and depth-based data minimization. AUC is reported for video anomaly detection on UCSD Ped1, UCSD Ped2, and CUHK Avenue, while privacy metrics cMAP and F1 are reported on PAHMDB. Higher VAD performance and lower privacy performance are considered better and represented in bold . The best optimal trade-off is represented in blue and

s

denotes the stride used in temporal subsampling.

Figure 2: Heatmap visualization of raw performance metrics and derived selection criteria for all single and pairwise minimization settings. The top row shows AUC, cMAP, and F1, while the bottom row reports distance to the ideal point, weighted score, and combined rank. Diagonal entries denote single minimization settings, and off-diagonal entries denote pairwise settings, where the minimization on the vertical axis is applied first, and the minimization on the horizontal axis is applied second. TS: Temporal sampling; DS: Downsample; MK: Masking; BL: Blurring; BR: Background removal.

Figure 3: Projection of all candidate settings onto the AUC-cMAP and AUC-F1 planes. Higher AUC and lower cMAP/F1 are preferred. Highlighted markers denote settings that are Pareto-optimal in the full three-objective space of AUC, cMAP, and F1. The figure provides a compact visualization of the feasible utility-privacy trade-off surface and complements the quantitative ranking reported in Section 4.4.2 .

Explore similar work

Privacy-Aware Video Anomaly Detection through Orthogonal Subspace Projection
May 9, 2026Lei Wang, Wenxiang Diao, Andrew Busch +2Privacy-Preserving ML Video Anomaly Detection
School of Engineering and Built Environment, Griffith University · Data61/CSIRO · School of Computer Science and Engineering, University of New South Wales +1
PA-VAD: Diffusion-Based Pseudo-Only Video Anomaly Detection via Domain-Aligned Memory Updates
Dec 7, 2025Satoshi Hashimoto, Yanan Wang, Hitoshi Nishimura +1Video Anomaly Detection Synthetic Anomaly Generation
KDDI Research, Inc., Fujimino, Saitama, Japan
From Frames to Events: Rethinking Evaluation in Human-Centric Video Anomaly Detection
Apr 10, 2026Narges Rashvand, Shanle Yao, Armin Danesh Pazho +2Video Anomaly Detection Temporal Action Detection
University of North Carolina at Charlotte, Charlotte, NC, USA

May 9, 2026J/K move · Enter open · S save

Privacy-Aware Video Anomaly Detection through Orthogonal Subspace Projection

Lei Wang, Wenxiang Diao, Andrew Busch, Jun Zhou+1

School of Engineering and Built Environment, Griffith University · Data61/CSIRO · School of Computer Science and Engineering, University of New South Wales+1

Privacy-Preserving ML Video Anomaly Detection Dimensionality Reduction

Open paper AskPDF arXiv

Video anomaly detection (VAD) systems often prioritize accuracy while overlooking privacy concerns, limiting their suitability for real-world deployment. We propose the Orthogonal Projection Layer (OPL), a lightweight module that removes task-irrelevant variations to produce representations focused on anomaly-relevant cues. To address privacy risks in human-centered scenarios, we introduce Guided OPL (G-OPL), which suppresses facial attributes using weak supervision from face-presence signals while preserving non-identifying features such as pose and motion. A cosine alignment objective enforces consistent capture and removal of facial information without identity labels or adversarial training. We further present a privacy-aware evaluation framework that jointly assesses detection performance and privacy preservation, and enables analysis of how sensitive information is filtered. Experiments show that embedding privacy constraints into model design reduces sensitive information while maintaining or improving detection accuracy, supporting projection-based architectures as a principled approach for privacy-aware VAD.

Abstract

Explore similar work

Privacy-Aware Video Anomaly Detection through Orthogonal Subspace Projection

PA-VAD: Diffusion-Based Pseudo-Only Video Anomaly Detection via Domain-Aligned Memory Updates

From Frames to Events: Rethinking Evaluation in Human-Centric Video Anomaly Detection