Data Anonymization

Blurring or removing personal information like faces, license plates, and voices from data so specific individuals can't be identified.

Data anonymization means removing or masking information that could identify a specific person before data is released or used for training. China's Personal Information Protection Law distinguishes two levels: de-identification means the data can no longer identify a specific natural person without additional information; anonymization means identification is impossible and irreversible, and anonymized information is no longer treated as personal information at all. Embodied-AI data increasingly comes from first-person video and crowdsourced collection in homes, shops, and factories, where footage captures bystanders' and operators' faces, house numbers, and screen content, and audio may capture conversations — so anonymization is a required step before releasing a dataset. The common approach is to use a detection model to automatically find and blur faces and license plates, remove or process audio, and then spot-check manually. Ego4D, for example, blurred the faces of bystanders and the license plates of passing cars in its videos, and stripped the audio from many clips entirely.

ExampleMeta open-sourced a model called EgoBlur for its Project Aria glasses, which automatically detects and blurs faces and license plates in first-person video; it's released under the Apache 2.0 license, which permits commercial use.

Also called

De-identification, Data De-identification

Related

Crowdsourced Data Collection · Egocentric Video · Ego4D · Project Aria Glasses · Data Cleaning · Data Quality Control

Sources

中华人民共和国个人信息保护法（中央网信办） (Chinese)
EGO4D's approach to privacy and ethics in data collection
EgoBlur（Project Aria）