West Virginia University 1
Removing Gender Signature from Fingerprints
Luca Lugini, Emanuela Marasco, Bojan Cukic, Jeremy Dawson
Lane Department of Computer Science and Electrical Engineering
Biometrics & Forensics & De-identification and Privacy Protection
May 29th 2014, Opatija
West Virginia University 2
Outline
• Privacy Issue
• Automatic gender estimation
• Gender de-identification
• Results
West Virginia University 3
ProblemPrivacy Protection
• Fingerprints may reveal valuable secondary information: age / gender, and other characteristics [1] [2].
• Possibly beneficial for forensic applications, but undesirable for most biometric use cases.
Goal:Obfuscation of age / gender through fingerprint image de-identification, without degrading match accuracy.
[1] E. Marasco, L. Lugini, and B. Cukic, “Exploiting Quality and Texture features to Estimate Age and Gender through Fingerprint Images”, SPIE
Defense and Security, pp. 1–10, 2014.
[2] P. Gnanasivam, and S. Muttan, “Fingerprint Gender Classification using Wavelet Transform and Singular Value Decomposition” International Journal
of Biometrics and Bioinformatics (IJBB) 9(2) (2012).
Related Work
West Virginia University 4
Literature on Gender Classification from Fingerprints
De-Identification Literature
• Manual Method (2006) [1]:Ridge count, Ridge thickness to valley thickness ratio, white lines count, pattern type concordanceDataset: 1100 males, 1100 femalesResults: GAR=88.28% Neural Network, GAR=86.5% LDA, GAR=80.39% Fuzzy C-Mean
•Automated Method (2012) [2]: Discrete Wavelet Transform (DWT), Singular Value Decomposition (SVD)Dataset: 1980 males, 1590 females; age groups: up to 12, 13-19, 20-25, 26-35, 36 and aboveResults: GAR=88.28%
For Faces: K-Same algorithm [3] & Extension of the K-Same algorithm [4]• From face similarities an algorithm creates new faces. • It guarantees privacy when sharing video data by preserving facial details but ensuring a low
reliability of face recognition• The extension takes into account linear appearance variations of faces
[1] Badawi, A., Mahfouz, M., Tadross, R., & Jantz, R. (2006, June). Fingerprint-Based Gender Classification. In IPCV (pp. 41-46)[2] Gnanasivam, P., S. Muttan. "Fingerprint Gender Classification using Wavelet Transform and Singular Value Decomposition”, IEEE-2012[3] E. Newton, L. Sweeney, B. Malin, “Preserving Privacy by De-Identifying Face Images,” IEEE Transactions on Knowledge and Data Engineering, vol. 17, no. 2, pp. 232–243, 2005. [4] R. Gross, L. Sweeney, J. Cohn, F. de la Torre, and S. Baker, “Face De-Identification,” Protecting Privacy in Video Surveillance, pp. 129–146, 2009.
5
The Proposed Approach
West Virginia University
• Ad-hoc image filtering and scaling in the frequency domain diminishes gender patterns.
• The distortion is introduced in both the gallery and probe fingerprints.
Automatic Gender Estimation
West Virginia University 6
E. Marasco, L. Lugini, and B. Cukic, “Exploiting Quality and Texture features to Estimate Age and Gender through Fingerprint Images”, SPIE Defense and Security, pp. 1–10, 2014.
• Impact of Gender on Matching and Image Quality
• NFIQ• Energy Concentration• Histogram of LBP• Histogram of LPQ• Entropy
• Energy Distributions• Gender Estimation Results
10-Fold Cross Validation
7
De-Identification Algorithm
West Virginia University
Input:Let I(x,y) be the original fingerprint image.Let B be the number of frequency bands.Let w0 and w1 represent the male and female classes, respectively.
Output:DI(x,y) de-identified image.
1. Compute the Discrete Fourier Transform F(u,v) of the image I(x,y).
MALE
FEMALE
8
De-Identification Algorithm
West Virginia University
2. Filter the image in the frequency domain: Gk(u,v) = Hk(u,v)F(u,v); for k = 1,…B
3. Estimate the energy distributions for w0 and w1: Ek,w0 = |Gk(u,v)|2
Ek,w1 = |Gk(u,v)|2; for k = 1,…B
4. Compute scaling parameters ak, for k = 1,…B µk,w0 = mean(Ek,w0) µk,w1 = mean(Ek,w1) ak = (µk,w0 - µk,w1)/2
5. Apply the scaling in the frequency domain: If w1 F*(u,v) = (ak + 1) F(u,v) else F*(u,v) = ak F(u,v)
6. Compute the inverse of F(u,v)
West Virginia University 9
Dataset • Data collection performed at West Virginia University
• FBI Certified livescan fingerprint systems
• Number of participants: 500
• Two sequential sessions of fingerprints for each sensor
• Rolled individual fingerprints on right and left hands; left, right and thumb slaps per session
– In the analysis we use right point finger only
West Virginia University 10
Visual Results
Female subject
before and after de-identification
Male subject
before and after de-identification
• Visually, the impact of de-identification process on the fingerprint images is not pronounced
11
Gender Estimation after De-Identification
West Virginia University
Energy distributions of specific frequency bands after de-identification
• Frequency components in original images separate females from males well• Corresponding frequency components in de-identified images for males and
females overlap• The initial gender estimation accuracy of 88.7% before de-identification is
reduced to 50.5%
Energy distributions of specific frequency bands before de-identification
Matching De-Identified Images
West Virginia University 12
Verification Performance before and after De-Identification
• Variations induced in the images do not drastically affect verification error rates, as expected from an effective de-identification algorithm
13
Conclusions
West Virginia University
• The initial gender estimation accuracy of 88.7% before de-identification is reduced to 50.5%
• Visually, the impact of de-identification process on the fingerprint images is not pronounced
• Variations induced in the images do not drastically affect verification error rates, as expected from an effective de-identification algorithm
• We propose a new de-identification algorithm to remove gender signature from fingerprints
• Automatic estimation of gender from fingerprints arises concerns about privacy protection
14West Virginia University
Any Questions?
Thanks for your Attention!
[email protected]: (304) 293-1455
Emanuela Marasco, Ph.D.WVU CITeR
Statler College of Engineering and Mineral ResourcesLCSEE – PO Box 6109
395 Evansdale Drive, ESB Annex 171 Morgantown WV 26506 USA