EgoVeil: Personalized Privacy Reasoning for Egocentric Video
Organizations: University of California, Los Angeles, Los Angeles, CA, USA
Abstract
Recent advances in vision-language models (VLMs) enable AI systems to interpret egocentric video. Yet continuous visual capture raises privacy concerns because sensitive content is pervasive and varies across users and contexts. We present EgoVeil, a personalized privacy reasoning system that determines which visual content to obscure based on individual preferences while retaining context useful for downstream assistance. It converts localized visual regions into structured semantic observations and combines a privacy reasoner with an explicit user-specific memory for personalized privacy reasoning. This design supports personalization and continual adaptation as preference evidence accumulates. We evaluate on in-the-wild data from 10 smart-glasses users under a cross-user chronological protocol. EgoVeil achieves 91.93% balanced accuracy and 84.42% Mask recall, outperforming direct VLM baselines and other controlled privacy reasoning baselines. Our results show that personalized, adaptive privacy reasoning better aligns privacy preservation with individual preferences while retaining the visual context needed for effective smart-glasses assistance.
Figures & tables
| Category | Example Questions | Answer | Proportion (%) |
| Object Localization | Where is my backpack? | Beside the chair. | 10.62 |
| Where did I put my phone? | On the dining table. | ||
| In what location did I see the shopping cart? | Grocery store. | ||
| Object Attributes | What did I put in the refrigerator? | A bottle of milk. | 47.39 |
| How many drinks were on the table? | Three. | ||
| What food did I place on the plate? | Sandwiches. |
| Field | Description | Example |
|---|---|---|
| Region Type | Visual entity represented by the region | Shipping Label |
| Visual State | Directly observable state of the region | Partially readable |
| Visible Information | Concrete information visibly conveyed by the region | Recipient name, home address |
| Scene Context | Immediate context surrounding the region | Package near a home entrance |
| Method | Bal. Accuracy (%) | Precision (%) | Recall (%) | F1 (%) |
|---|---|---|---|---|
| General VLM baselines | ||||
| Generic VLM | 53.85 | 43.02 | 7.93 | 7.29 |
| Generic VLM + Memory | 62.10 | 17.60 | 25.35 | 19.36 |
| Controlled reasoning ablations | ||||
| Base Privacy Reasoner | 74.59 | 15.65 | 60.17 | 19.24 |
| Base Privacy Reasoner + Memory | 85.23 | 24.34 | 73.99 | 31.37 |
| Video Condition | Qwen3-VL-8B (%) | InternVL3.5-8B (%) |
|---|---|---|
| Raw Video | 49.33 24.98 | 55.33 22.67 |
| Obfuscated Video | 42.67 24.38 | 52.00 19.06 |
Privacy Protection in VLMs
Privacy protection in vision-language models limits sensitive information exposure while preserving useful image understanding and multimodal reasoning.
See all 44 papers