1. 1
Applying Transform-domain Scrambling
to Automatically Detected Faces
Pavel Korshunov, Aleksei Triastcyn, and
Touradj Ebrahimi
Multimedia Signal Processing Group
Swiss Federal Institute of Technology, Lausanne
2. Introduction 2
•Recipe
– Take simple face detection
(OpenCV)
– Combine with a privacy filter
– Apply filter to all detected regions
in a video
•Focus on the privacy filter
Multimedia Signal Processing Group
Swiss Federal Institute of Technology, Lausanne
3. Privacy Filters 3
• Simple: pixelization, masking, and
blurring
– Non-reversible
– encryption anonymization, etc.
• Anonymization (replacing with
another object)
– Non-reversible
– Hard to implement
• Encryption
– Video alterations break the filter
– Complex
Multimedia Signal Processing Group
Swiss Federal Institute of Technology, Lausanne
4. Transform-domain scrambling 4
Scramblin Entropy
frame Transform
g coding
bitstream
encoder
• Seed random generator with a secret key
• Randomly flip sign of 63 DCT coefficients
in a scrambled macro block
• During decoding, repeat the same
Multimedia Signal Processing Group
Swiss Federal Institute of Technology, Lausanne
5. Scrambling in JPEG 5
F. Dufaux and T. Ebrahimi, “Scrambling for privacy protection in video surveillance systems,” IEEE
Trans. on Circuits and Systems for Video Technology, vol. 18, no. 8, pp. 1168–1174, Aug 2008.
Multimedia Signal Processing Group
Swiss Federal Institute of Technology, Lausanne
6. Transform-domain scrambling 6
• Pros
– Reversible method
– Does not negatively affect coding efficiency
– Scrambling strength can be controlled
– Security can be insured
• Cons
– Must be integrated inside the encoder
Multimedia Signal Processing Group
Swiss Federal Institute of Technology, Lausanne
7. Subjective evaluation results 7
Detection
accuracy:
0.24
Multimedia Signal Processing Group
Swiss Federal Institute of Technology, Lausanne
9. Scrambled frame, morning video 9
Multimedia Signal Processing Group
Swiss Federal Institute of Technology, Lausanne
10. Scrambled frame, evening video 10
Multimedia Signal Processing Group
Swiss Federal Institute of Technology, Lausanne
11. Scrambled frame, evening video 11
Multimedia Signal Processing Group
Swiss Federal Institute of Technology, Lausanne
12. Conclusion 12
•Face detection played unexpectedly
important role
– Either use better detection or re-think
evaluation methodology
•Subjects were highly irritated with
scrambling!
– Make scrambling more human-friendly
Multimedia Signal Processing Group
Swiss Federal Institute of Technology, Lausanne