Low-Latency Video Anonymization for Crowd Anomaly Detection: Privacy Versus Performance

Technical Summary: Low-Latency Video Anonymization for Crowd Anomaly Detection

Problem Formulation
The proliferation of surveillance cameras and AI-driven monitoring systems presents a critical tension between public safety and individual privacy. While deep learning (DL) models excel at Video Anomaly Detection (VAD) for crowd monitoring, they often rely on sensitive biometric data, raising concerns regarding privacy violations, model bias, and regulatory non-compliance (e.g., GDPR, EU AI Act). Existing solutions face a dichotomy:

1. Conventional Anonymization (AN): Techniques like masking, pixelization, and blurring are computationally lightweight and suitable for real-time edge deployment but often use fixed hyperparameters. This leads to inadequate privacy protection for subjects at varying depths or sizes, or conversely, excessive image degradation that compromises VAD utility.
2. Deep Learning-Based AN: Generative models (e.g., inpainting) offer high-quality anonymization but are computationally expensive, making them unsuitable for resource-constrained edge devices. Furthermore, many DL-based approaches require retraining downstream VAD models on anonymized data.

The core challenge addressed is achieving a balance where privacy protection is robust enough to prevent biometric leakage (face, gender, ethnicity) without significantly degrading the utility of the video for downstream anomaly detection tasks, all while maintaining low latency for real-time edge deployment.

Methodology
The authors propose LA3D (Lightweight Adaptive Anonymization for VAD), a system designed to dynamically adjust conventional anonymization parameters based on the visual characteristics of the target subjects.

• Adaptive Mechanism: Unlike static approaches that apply a uniform blur or pixelation factor to an entire frame, LA3D first segments human subjects using a lightweight object detection model (YOLOv8-Seg). It then calculates an adaptive scaling factor ($r_i$) for each detected subject based on the ratio of the subject's surface area to the total image area.
• Dynamic Parameter Adjustment:
  • Gaussian Blurring ($G$): The kernel size ($k$) and standard deviation ($\sigma$) are scaled according to the subject's size. Larger subjects receive stronger blurring to ensure privacy, while smaller subjects are treated proportionally to avoid unnecessary image degradation.
  • Pixelization ($P$): The downsampling factor ($d$) is similarly adjusted. The system includes a "Maximum Adaptive" mode ($AN_{max}=1$) where the entire subject region is downsampled to a single pixel (creating a silhouette) if the adaptive logic dictates maximum protection.
• Privacy Attribute Detection (PD) Model: To evaluate privacy leakage, the authors developed a multi-label classifier (ResNet50) trained on the VISPR dataset. This model detects six privacy attributes: Gender, Face, Nudity, Skin Color, Relationship, and "No Attribute." The model utilizes a pre-trained encoder and a class-weighted loss function to address class imbalance in the training data.
• VAD Utility Evaluation: The utility of the anonymized video is assessed using state-of-the-art Weakly Supervised VAD (WSAD) models, specifically PEL4VAD and MGFN, on the UCF-Crime and XD-Violence datasets. These models were trained on raw (non-anonymized) data and tested on anonymized data to measure performance retention without retraining.

Key Contributions

1. Revisiting Conventional AN: The study demonstrates that optimized conventional algorithms can outperform computationally heavy DL-based anonymization methods in terms of efficiency and effectiveness for edge deployment.
2. Adaptive Full-Body Anonymization: The introduction of LA3D, which dynamically tunes hyperparameters based on subject depth and size, significantly improves the trade-off between privacy and utility compared to fixed-parameter baselines.
3. Comprehensive Benchmarking: The paper provides a detailed performance report across three dimensions:
   • Privacy: Leakage detection via PD and Person Re-Identification (ReID) on the Market1501 dataset.
   • Utility: Anomaly detection accuracy (AUC/AP) on UCF-Crime and XD-Violence.
   • Efficiency: Computational cost analysis comparing CPU/GPU processing times and memory usage against DL-based alternatives.
4. Improved PD Model: An enhanced privacy attribute detector augmented with pre-trained encoders and weighted loss functions to better handle imbalanced privacy classes.

Results

• Privacy Protection: LA3D significantly reduces privacy attribute leakage. On the VISPR dataset, adaptive methods achieved a 27.35%–40.40% reduction in Mean Average Precision (mAP) for privacy detection compared to non-anonymized data, outperforming non-adaptive baselines. In ReID tasks on Market1501, adaptive pixelization ($P^a$) reduced the mean Average Precision (mAP) by over 95%, effectively preventing person re-identification.
• VAD Utility: The adaptive approach maintains high VAD efficacy. On the UCF-Crime dataset, the adaptive Gaussian blurring ($G^a$) actually improved VAD performance slightly (0.12%–0.23% gain) compared to non-adaptive blurring, likely due to noise reduction. Even for pixelization, the utility degradation was minimal (often <1% for bounded adaptive settings) compared to the significant privacy gains.
• Computational Efficiency: LA3D is highly efficient. The adaptive processing adds only 5–6% overhead to the base non-adaptive algorithms. In contrast, DL-based approaches (e.g., DeepPrivacy2) are approximately 16 times slower and consume 14 times more memory. LA3D is feasible for real-time edge deployment, whereas DL-based methods struggle with resource constraints.

Significance and Claims
The paper claims that lightweight, adaptive conventional algorithms offer a superior alternative to both static conventional methods and computationally expensive deep learning approaches for privacy-preserving VAD. The significance lies in demonstrating that:

1. Dynamic adaptation is crucial: Fixed parameters fail to handle the variability of real-world surveillance footage (varying depths and subject sizes), leading to either privacy leaks or unnecessary utility loss.
2. Privacy and Utility are not strictly trade-offs: With the right adaptive mechanism, it is possible to achieve substantial privacy protection (outperforming DL-based methods in accuracy and efficiency) without severely degrading VAD performance.
3. Edge Feasibility: The proposed method enables privacy-aware, real-time video monitoring on edge devices, addressing the growing need for compliant and efficient crowd anomaly detection systems without the infrastructure costs of cloud-based processing or the latency of generative AI.

The authors conclude that their approach provides a robust baseline for future privacy-related studies and highlights the potential of body-level, adaptive anonymization for practical, privacy-aware computer vision applications.