Workflow
Design Goals
Authors conducted an online survey to support their design goals
- Privacy policies should be individualized
- Privacy policies should be situational
- Compliance by courtesy is sufficient
Techniques
- Face detection: HeadHunter (high recall, low precision)
- Face recognition: DNN based, Imagenet finetuned feature extraction (128 dimension) + SVM
- Secure Matching Protocol: secure dot product + gabled circuits
Evaluation
15 volunteers carried ipic compliant devices with specified policies of when and where to show or blur their faces
Faces Collected
Result
Methodology of Analysis
- Analyzed visual pipeline failure by categorizing 1000+ collected images into different categories (good lighting, occluded…) and gives recall rate for each category.
- Mapped categories above into social events
- Compared authors’ face detection with others on recall rate.
- Analyzed computation, bandwidth, and energy cost of secure matching protocol