Can we still avoid automatic face detection?

WACV, 2016.

Cited by: 15|Bibtex|Views74
EI
Other Links: dblp.uni-trier.de|academic.microsoft.com|arxiv.org
Weibo:
Though we only evaluated Facebook’s face detector, our results could apply to any social medium that uses automatic face recognition technology

Abstract:

After decades of study, automatic face detection and recognition systems are now accurate and widespread. Naturally, this means users who wish to avoid automatic recognition are becoming less able to do so. Where do we stand in this cat-and-mouse race? We currently live in a society where everyone carries a camera in their pocket. Many pe...More

Code:

Data:

0
Introduction
  • Consider the “Social Network” setting in which a user takes a picture of their family having lunch in a park.
  • After uploading this picture, the social network automatically tags the people that appear in the user’s picture.
  • Depending on the user’s settings, it may automatically become searchable by date, time, place, or the name of the people who appear in the photo.
  • How does this complicated process work? Traditional recognition systems typically use five steps:
Highlights
  • The results reveal that state-of-the-art face detectors generally have trouble recognizing faces that are severely out-of-focus or small
  • There is still a large gap in performance—human workers achieve 60% face detection accuracy when as little as 25% of the face is visible, but Picasa requires 60% of the face to be visible for similar levels of accuracy
  • Though we only evaluated Facebook’s face detector, our results could apply to any social medium that uses automatic face recognition technology
  • After all, avoiding automatic face detection and recognition is becoming more difficult as talented engineers search for ways to improve these systems
  • The photo uploader, not the individual, must remember to use the image perturbation techniques. Many of these techniques make the image look worse to humans. These techniques illuminate the strengths and weaknesses of state-of-the-art face detectors used in common social network platforms
Results
  • There is still a large gap in performance—human workers achieve 60% face detection accuracy when as little as 25% of the face is visible, but Picasa requires 60% of the face to be visible for similar levels of accuracy.
  • An image is considered a true accept if any of Facebook’s boxes have an Intersection-over-Union (IoU) score greater than 10% with the groundtruth.1
Conclusion
  • Though the authors only evaluated Facebook’s face detector, the results could apply to any social medium that uses automatic face recognition technology.
  • After all, avoiding automatic face detection and recognition is becoming more difficult as talented engineers search for ways to improve these systems.
  • The photo uploader, not the individual, must remember to use the image perturbation techniques
  • Many of these techniques make the image look worse to humans.
  • If a privacy-seeking individual wishes to develop more ways of avoiding automatic detection, building from these observations could be a good first start
Summary
  • Introduction:

    Consider the “Social Network” setting in which a user takes a picture of their family having lunch in a park.
  • After uploading this picture, the social network automatically tags the people that appear in the user’s picture.
  • Depending on the user’s settings, it may automatically become searchable by date, time, place, or the name of the people who appear in the photo.
  • How does this complicated process work? Traditional recognition systems typically use five steps:
  • Objectives:

    The authors' goal is not to complain about Facebook’s default policy settings or lament the impending death of privacy.
  • Results:

    There is still a large gap in performance—human workers achieve 60% face detection accuracy when as little as 25% of the face is visible, but Picasa requires 60% of the face to be visible for similar levels of accuracy.
  • An image is considered a true accept if any of Facebook’s boxes have an Intersection-over-Union (IoU) score greater than 10% with the groundtruth.1
  • Conclusion:

    Though the authors only evaluated Facebook’s face detector, the results could apply to any social medium that uses automatic face recognition technology.
  • After all, avoiding automatic face detection and recognition is becoming more difficult as talented engineers search for ways to improve these systems.
  • The photo uploader, not the individual, must remember to use the image perturbation techniques
  • Many of these techniques make the image look worse to humans.
  • If a privacy-seeking individual wishes to develop more ways of avoiding automatic detection, building from these observations could be a good first start
Tables
  • Table1: Summary of our filters, approximate detection accuracies under the strongest settings, and subjective amount of image degradation
  • Table2: Facebook’s detection probability on two datasets of face images with naturally-occluding clothing
Download tables as Excel
Related work
  • Many projects consider privacy enhancement tools that foil face detection by altering the actual face’s appearance as seen by the imaging sensor. For example, consider the Privacy Visor designed by Echizen et al [24]. This device consists of several high-powered infrared LEDs mounted on a pair of glasses. When activated, the infrared LEDs wash out the face when captured with conventional digital cameras, but are invisible to the human eye. The user’s friends see a person wearing funny glasses, but their cameras only see specular highlights. Another similar project is Adam Harvey’s CV Dazzle [9], which is a fashion statement as much as it is a privacy tool. Rather than wearing glasses, users can wear make-up or bits of plastic to foil face detection in a stylish, eye-catching way.
Funding
  • There is still a large gap in performance—human workers achieve 60% face detection accuracy when as little as 25% of the face is visible, but Picasa requires 60% of the face to be visible for similar levels of accuracy
  • An image is considered a true accept if any of Facebook’s boxes have an Intersection-over-Union (IoU) score greater than 10% with the groundtruth.1
Study subjects and analysis
subjects: 135
Example images are shown in Fig. 7. The classic ARFace dataset [13], now 18 years old, contains carefully captured pictures of 135 subjects obtained under a controlled laboratory setting. For our setting, we use 417 “neutral” images (subsets 1, 5, and 6), 417 images of subjects wearing scarves (subsets 11, 12, 13), and 417 images of subjects with sun glasses (subsets 8, 9, and 10)

images of subjects: 417
The classic ARFace dataset [13], now 18 years old, contains carefully captured pictures of 135 subjects obtained under a controlled laboratory setting. For our setting, we use 417 “neutral” images (subsets 1, 5, and 6), 417 images of subjects wearing scarves (subsets 11, 12, 13), and 417 images of subjects with sun glasses (subsets 8, 9, and 10). We also evaluate performance on the UMB-DB 3D face dataset [7], which contains 2D and 3D captures of 143 subjects

subjects: 143
For our setting, we use 417 “neutral” images (subsets 1, 5, and 6), 417 images of subjects wearing scarves (subsets 11, 12, 13), and 417 images of subjects with sun glasses (subsets 8, 9, and 10). We also evaluate performance on the UMB-DB 3D face dataset [7], which contains 2D and 3D captures of 143 subjects. Each image is tagged with several binary attributes such as “scarf,” “smile,” “free,” and “occluded.”

Reference
  • B. Babenko, M.-H. Yang, and S. Belongie. Robust object tracking with online multiple instance learning. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 33(8):1619–1632, 2011.
    Google ScholarLocate open access versionFindings
  • P. Belhumeur, D. Jacobs, D. Kriegman, and N. Kumar. Localizing parts of faces using a consensus of exemplars. In CVPR, pages 545–552, June 2011.
    Google ScholarLocate open access versionFindings
  • T. Berg and P. N. Belhumeur. Tom-vs-pete classifiers and identity-preserving alignment for face verification. In BMVC, 2012.
    Google ScholarLocate open access versionFindings
  • V. Bruce, M. A. Burton, and N. Dench. What’s distinctive about a distinctive face? The Quarterly Journal of Experimental Psychology Section A, 47(1):119–141, Feb. 1994.
    Google ScholarLocate open access versionFindings
  • X. Burgos-Artizzu, P. Perona, and P. Dollar. Robust Face Landmark Estimation under Occlusion. In ICCV Workshops, pages 1513–1520, Dec. 2013.
    Google ScholarLocate open access versionFindings
  • X. P. Burgos-Artizzu, P. Perona, and P. Dollr. Robust face landmark estimation under occlusion. In ICCV, 2013.
    Google ScholarLocate open access versionFindings
  • A. Colombo, C. Cusano, and R. Schettini. UMB-DB: A database of partially occluded 3d faces. In ICCV Workshops, pages 2113–2119, Nov. 2011.
    Google ScholarLocate open access versionFindings
  • D. Goldman. Facebook is scanning your photos to help you share them. CNN Money, Nov. 2015.
    Google ScholarLocate open access versionFindings
  • A. Harvey. CV Dazzle: Camouflage from computer vision. Technical report, New York University, 2012.
    Google ScholarFindings
  • G. B. Huang, M. Ramesh, T. Berg, and E. LearnedMiller. LFW results. Internet: http://vis-www.cs.umass.edu/lfw/results.html.[Accessed: Sep.25, 2015].
    Findings
  • F. Juefei-Xu, D. Pal, K. Singh, and M. Savvides. A Preliminary Investigation on the Sensitivity of COTS Face Recognition Systems to Forensic Analyst-Style Face Processing for Occlusions. In CVPR, pages 25– 33, 2015.
    Google ScholarLocate open access versionFindings
  • N. Kumar, A. C. Berg, P. N. Belhumeur, and S. K. Nayar. Attribute and Simile Classifiers for Face Verification. In ICCV, Oct 2009.
    Google ScholarLocate open access versionFindings
  • A. M. Martinez and R. Benavente. The AR Face database. CVC Technical Report, 24, 1998.
    Google ScholarFindings
  • J. Parris, M. Wilber, B. Heflin, H. Rara, A. ElBarkouky, A. Farag, J. Movellan, M. CastrilonSantana, J. Lorenzo-Navarro, M. Teli, S. Marcel, C. Atanasoaei, and T. Boult. Face and eye detection on hard datasets. In 2011 International Joint Conference on Biometrics (IJCB), 2011.
    Google ScholarLocate open access versionFindings
  • S. Ren, X. Cao, Y. Wei, and J. Sun. Face alignment at 3000 FPS via regressing local binary features. In CVPR, pages 1685–1692. IEEE, 2014.
    Google ScholarLocate open access versionFindings
  • J. A. Sarno and T. R. Alley. Attractiveness and the memorability of faces: Only a matter of distinctiveness? The American Journal of Psychology, 110(1):81–92, Apr. 1997.
    Google ScholarLocate open access versionFindings
  • W. Scheirer and T. Boult. Biometrics: Practical issues in privacy and security. Technical report, Securics, Inc. and the University of Colorado Colorado Springs, 2011. Tutorial at International Jount Conference on Biometrics.
    Google ScholarFindings
  • W. J. Scheirer, S. E. Anthony, K. Nakayama, and D. D. Cox. Perceptual annotation: Measuring human vision to improve computer vision. IEEE Transactions on Pattern Analysis and Machine Intelligence (T-PAMI), 36, August 2014.
    Google ScholarLocate open access versionFindings
  • Y. Sun, X. Wang, and X. Tang. Deep learning face representation by joint identification-verification. arXiv:1406.4773 [cs], June 2014. arXiv: 1406.4773.
    Findings
  • Y. Taigman, M. Yang, M. Ranzato, and L. Wolf. Deepface: Closing the gap to human-level performance in face verification. In CVPR, 2013.
    Google ScholarFindings
  • P. Viola and M. Jones. Robust real-time face detection. International Journal of Computer Vision (IJCV), 57(2):137–154, May 2004.
    Google ScholarLocate open access versionFindings
  • H. Wang and J. Chen. Improving self-quotient image method of NPR. In 2008 International Conference on Computer Science and Software Engineering, volume 6, pages 213–216, 2008.
    Google ScholarLocate open access versionFindings
  • H. Wang, S. Li, and Y. Wang. Face recognition under varying lighting conditions using self quotient image. In Sixth IEEE International Conference on Automatic Face and Gesture Recognition, 2004. Proceedings, pages 819–824, 2004.
    Google ScholarLocate open access versionFindings
  • T. Yamada, S. Gohshi, and I. Echizen. Privacy Visor: Method for Preventing Face Image Detection by Using Differences in Human and Device Sensitivity. In B. D. Decker, J. Dittmann, C. Kraetzer, and C. Vielhauer, editors, Communications and Multimedia Security, number 8099 in Lecture Notes in Computer Science, pages 152–161. Springer Berlin Heidelberg, Sept. 2013. DOI: 10.1007/978-3-642-40779-6 13.
    Locate open access versionFindings
Full Text
Your rating :
0

 

Tags
Comments