Research map / Video / Action recognition
Selective Interpretable and Motion Consistent Privacy Attribute Obfuscation for Action Recognition
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a training-free privacy obfuscation module for action recognition that uses human-selected DINO-ViT template matching to build selective saliency maps and applies optical-flow-warped noise to hide privacy attributes while preserving motion and context.
Problem
Privacy-preserving action recognition suffers from global obfuscation that hides action-relevant context, loss of motion information, need for retraining the recognition models, and lack of interpretability.
Approach
Selective, interpretable obfuscation: DINO-ViT key-based template matching from a human-curated library of anatomical/identity templates yields saliency maps that weight temporally consistent noise warped by RAFT optical flow, applied directly to input video without retraining any action or privacy network.
Key results
- Proposed method reaches f0.5 of 0.68 on IPN, 0.92 on KTH and 0.87 on SBU, beating the best naive baseline and SOTA methods on all three datasets.
- Action accuracy drops at most 1.48 points (KTH 90.15 to 88.67) while privacy accuracy falls sharply, e.g. KTH privacy 91.87 to 5.46.
- Temporally consistent noise outperforms iid noise for action recognition on all datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use selective, motion-consistent template-guided obfuscation instead of BDQ | Action and privacy recognition · IPN | f0.5: 0.61 -> 0.68 | Helped |
| Use selective, motion-consistent template-guided obfuscation instead of BDQ | Action and privacy recognition · KTH | f0.5: 0.92 -> 0.92 | No clear effect |
| Use selective, motion-consistent template-guided obfuscation instead of ELR s=16 | Action and privacy recognition · SBU | f0.5: 0.57 -> 0.87 | Helped |
| Replace motion-consistent noise with iid noise (ablation) | Ours · IPN, KTH, and SBU | Action accuracy | Hurt |
| Replace motion-consistent noise with iid noise (ablation) | Ours · IPN, KTH, and SBU | Privacy accuracy | No clear effect |
| Obfuscate the hand template (ablation) | Ours · IPN | Action accuracy | Hurt |
| Select dataset-specific privacy templates rather than obfuscating all templates (ablation) | Ours · IPN, KTH, and SBU | Action accuracy | Mixed |
Surprising findings
- Pixelation 4x4 performs strongly on IPN and KTH among naive baselines, showing that simple global methods can be competitive on the trade-off metric.
Limitations
- Trade-off between action and privacy since motion preserved may enable gait-based identification.
- Template library focused on identity attributes; generalization to other attributes requires manual template design.
- Comparisons to SOTA methods limited to a single recognition network in some cases.
Experimental setup
Datasets: IPN SBU KTH Kinetics400 ImageNet-1k
Benchmarks: IPN KTH SBU
Metrics: top-1 accuracy f_lambda trade-off score
Baselines: ELR BDQ ALF Pixelation Blur Masking (YOLOv8)
Builds on: DINO RAFT Visual correspondence via ViT keys Motion-preserving privacy obfuscation (Ellis/E2S)
Closest papers by meaning
- Joint Attribute and Model Generalization Learning for Privacy-Preserving Action Recognition (NeurIPS 2023 · similarity 0.87)
- MaSS: Multi-attribute Selective Suppression for Utility-preserving Data Transformation from an Information-theoretic Perspective (ICML 2024 · similarity 0.83)
- MultiPriv: Benchmarking Individual-Level Privacy Reasoning in Vision-Language Models (ICML 2026 · similarity 0.83)
- Private Attribute Inference from Images with Vision-Language Models (NeurIPS 2024 · similarity 0.82)
- PerceptAnon: Exploring the Human Perception of Image Anonymization Beyond Pseudonymization for GDPR (ICML 2024 · similarity 0.82)
- Learning Robust and Privacy-Preserving Representations via Information Theory (AAAI 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.