WIP: Blurred Lines -- A GDPR-Compliant Framework for Anonymising Automotive...
summary
A framework called Blurred Lines processes vehicle camera feeds with five detectors (Haar/cascade, YOLO, a face-specific model, OCR for plates, and a vehicle-region fallback) and multiple anonymization modes (Gaussian blur, pixelation, black boxes, face swapping). Tested on an industry dataset of over 50 hours, the pipeline runs as a command-line interface, returns near 100% accuracy on PII detection by favoring conservative over-blurring, and is already adopted by an industrial partner though heavy masking sometimes degrades ML utility.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:15Sensitive data in automotive videos
- 3:02Data sharing for AI development
- 5:09GDPR challenges in video anonymization
- 7:17Personally identifiable info in videos
- 8:40Blurred Lines framework
- 10:25Detecting and blurring faces
- 11:50Five anonymization methods
- 13:15Anonymization versus data utility
- 14:42Privacy versus ML usability
- 15:44Industry case study in Birmingham
- 17:09Tool availability and adoption
- 18:31Lessons learned and future work
- 19:33Conclusion
related talks
21mCombining Privacy and AI
19mGDPR, AI & use of Video data
32mPrivacy Protection. Privacy by Design: Strengthening Compliance
40mNew Face, Who Dis? Protecting Privacy in a World of Surveillance
01h 05mFacial recognition, a 'convenient' and 'efficient' solution looking for a problem?
01h 11mPRIVACY IN AUTOMATED AND CONNECTED VEHICLES
01h 15mHOW TO RECONCILE FACIAL RECOGNITION TECHNOLOGIES WITH CONSUMERS' PRIVACY
40mGDPR in practise - Developing models with transparency and privacy in mind