Research map / Federated learning and privacy / Privacy attacks and secure learning
EgoPrivacy: What Your First-Person Camera Says About You?
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.
Builds EgoPrivacy, a benchmark of paired ego-exo videos for measuring privacy leakage about the camera wearer (demographic, identity, situational), evaluates foundation-model attacks, and proposes a retrieval-augmented attack (RAA) that boosts demographic inference by retrieving exocentric views of the wearer.
Problem
Wearable cameras capture egocentric video that can reveal private information about the camera wearer themselves (demographics, identity, location/time), a threat that prior egocentric privacy work largely overlooked.
Approach
Frames privacy risk to the camera wearer across three categories and seven tasks, and introduces retrieval-augmented attack: ego-to-exo retrieval from an external pool of exocentric videos, followed by voting over the retrieved exo views plus the ego prediction.
Key results
- Zero-shot foundation models reach up to 73.15% gender, 65.36% race and 79.64% age accuracy, well above chance.
- RAA with VideoMAE improves gender accuracy by over 16 points; soft-voting RAA with VideoLLaMA2-7B raises race accuracy from 53.97% to 66.45%.
- Fine-tuned ego-to-exo identity retrieval with EgoVLP v2 reaches 50.31% HR@1 and ego-to-ego 81.25% HR@1 on ID test.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add retrieval-augmented attack to VideoLLaMA2-7B | VideoLLaMA2-7B demographic attack · Charades-Ego (OOD), gender | accuracy: +6.33 | Helped |
| Add retrieval-augmented attack to EgoVLP v2 | EgoVLP v2 demographic attack · Charades-Ego (OOD), race | accuracy: +3.56 | Helped |
| Add retrieval-augmented attack to EgoVLP v2 | EgoVLP v2 demographic attack · Charades-Ego (OOD), age | accuracy: +1.16 | Helped |
| Use hard voting instead of direct ego prediction (ablation) | VideoLLaMA2-7B demographic attack · Charades-Ego (OOD), gender | accuracy: 73.15 -> 78.32 | Helped |
| Use hard voting instead of direct ego prediction (ablation) | VideoLLaMA2-7B demographic attack · Charades-Ego (OOD), race | accuracy: 53.97 -> 65.82 | Helped |
| Use hard voting instead of direct ego prediction (ablation) | VideoLLaMA2-7B demographic attack · Charades-Ego (OOD), age | accuracy: 47.08 -> 43.23 | Hurt |
| Use soft voting with w=0.5 instead of direct ego prediction (ablation) | VideoLLaMA2-7B demographic attack · Charades-Ego (OOD), age | accuracy: 47.08 -> 51.12 | Helped |
| Use soft voting with w=1/(k+1) instead of direct ego prediction (ablation) | VideoLLaMA2-7B demographic attack · Charades-Ego (OOD), race | accuracy: 53.97 -> 66.45 | Helped |
| Add retrieval augmentation using soft voting | Demographic privacy attacks · Demographic tasks across evaluated models | accuracy | Mixed |
Surprising findings
- Zero-shot foundation models attack demographic privacy nearly as well as fine-tuned ones, implying minimal attacker capability suffices.
- Adding identity-level aggregation (capability 4) surprisingly lowers performance on ego age and race attacks.
Limitations
- Demographic labels are crowd-sourced perceived attributes from Amazon Mechanical Turk, not self-reported.
- Race restricted to Asian, Black, White; gender binary; dataset inherits imbalances from Ego-Exo4D and Charades-Ego.
- Moment-retrieval labels are approximated from clip footage rather than ground truth.
Experimental setup
Datasets: Ego-Exo4D Charades-Ego 11K Hands FairFace
Benchmarks: EgoPrivacy (Ego-Exo4D ID test) EgoPrivacy (Charades-Ego OOD test)
Metrics: accuracy HR@1 HR@5
Baselines: CLIP ViT-H/14 zero-shot EgoVLP v2 VideoMAE LLaVA-1.5 VideoLLaMA2 face-based demographic classifier (FairFace) hand-based demographic classifier (ResNet50)
Builds on: Ego-Exo4D Charades-Ego SupCon EgoVLP v2 VideoMAE
Closest papers by meaning
- PerceptAnon: Exploring the Human Perception of Image Anonymization Beyond Pseudonymization for GDPR (ICML 2024 · similarity 0.86)
- MultiPriv: Benchmarking Individual-Level Privacy Reasoning in Vision-Language Models (ICML 2026 · similarity 0.85)
- PID: Prompt-Independent Data Protection Against Latent Diffusion Models (ICML 2024 · similarity 0.83)
- Assessing Per-Sample Membership Inference Vulnerability without Retraining (NeurIPS 2026 · similarity 0.83)
- Fusing Personal and Environmental Cues for Identification and Segmentation of First-Person Camera Wearers in Third-Person Views (CVPR 2024 · similarity 0.83)
- Private Attribute Inference from Images with Vision-Language Models (NeurIPS 2024 · 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.