1. 1 Extracting Discriminative Binary Template for Face Template Protection Feng Yicheng Supervisor: Prof. Yuen August 31 st, 2009

2. 2 Content 1. Introduction 2. Basic Idea 3. Thresholding to Approximation 4. Objective Function Construction 5. Experimental Results 6. Conclusions

3. 3 Introduction Biometric for personal authentication has been used in many applications. Since biometric is the “unique” feature, it is hard to reset or re-issue. Security and privacy concern  Non-invertible: The attacker can’t extract the original templates with the data stored in database.  Cancelable: If some templates are compromised, new templates can be generated to replace them.  Application-specific: Different applications should use different versions of templates.

4. 4 Introduction Biometric cryptosystem approach is applied for protection  Require binary input Existing approaches apply thresholding to binarize the original biometric templates  Discriminability may be affected with the binarization  Effect to discriminability has not been evaluated Objective:  Find an approach to discriminatively binarize the face templates

5. 5 Basic Idea Use thresholding for binarization Directly optimizing thresholds has some problems  Contradict to the max-entropy rule Max-entropy rule: to gain maximum information content, the thresholds should be set to make half of the transformed bits to be 1, half to be 0.  Not effective Thresholds satisfying the max-entropy rule provides highest information content, already implying certain discriminability (Figure 1). Optimizing thresholds may not fit the data distribution (Figure 2).

6. 6 Basic Idea “Mean”: the thresholds are set as the mean values of the original templates. “Random”: thresholds are randomly chosen with a Gaussian distribution  Mean of the distribution is mean of the original templates  Variance of the distribution is r times of the variance of the original templates.  Tested 100 times, choose the average. Figure 1

7. 7 Basic Idea 2-dimensional scenario for thresholds optimization y x 0 threshold t 2 threshold t 1 Figure 2

8. 8 Basic Idea To fit the data distribution better, choose a projection before threhsolding  First do an projection, then do thresholding.  Fit data distribution better (Figure 3)  The projection should not degrade the discriminability: choose orthonormal matrix. The projection is discriminability preserving. Projection Original face template p MTpMTp Thresholding Binary template w

9. 9 Basic Idea Projection can make the thresholding fit the data distribution better. x 0 threshold t 2 threshold t 1 x 0 threshold t 2 threshold t 1 Projection Figure 3

10. 10 Basic Idea Proposed scheme  Original template p is first projected with orthonormal matrix M: u=M T p  u=(g 1, g 2 … g k ) is then thresholded to binary template (b 1, b 2 … b k ) with thresholds t 1, t 2 … t k. Due to the max- entropy rule, t i should be the mean value of g i.  Find optimal M to maximize the discriminability of the extracted binary templates.  For different classes, we choose different M.

11. 11 Basic Idea Discriminability measurement (for class Ω):  Within-class variance D W (Ω)  Between-class variance D B (Ω)  Discriminability: D B (Ω)- D W (Ω)  Optimization: w(p): the binary template transfromed from p. w Ω : the reference binary template of class Ω.

12. 12 Thresholding to Approximation Normalize p to simplify the thresholding Assume v=(a 1, a 2 … a k ), then the thresholding process turns to is the mean vector of all p.

13. 13 Thresholding to Approximation This process is equivalent to: Substitute v=M T q to this equation, w’(v) turns to subject to Replace the original thresholding

14. 14 Objective Function Construction M is orthonormal

15. 15 Objective Function Construction We can use D’ B (Ω) and D’ W (Ω) to replace D B (Ω) and D W (Ω). Denote. subject to q Ω represents the mean vector of q in class Ω. (distance from e to q in class Ω is small)

16. 16 Experimental Results Experiment settings  Three common face databases used CMU PIE (68x105x10) FERET (250x4x2) FRGC (350x40x5)  Fisherface algorithm applied for feature extraction  Compared with the RMQ algorithm

17. 17 Experimental Results CMU PIE

18. 18 Experimental Results FERET

19. 19 Experimental Results FRGC

20. 20 Experimental Results The GARs (FAR=0.01) and Equal Error Rates (EERs). GAROriginalRMQTOP CMU PIE59.2673.5385.99 FERET45.4774.0185.09 FRGC26.2867.4078.35 EEROriginalRMQTOP CMU PIE17.3210.376.30 FERET21.6611.297.44 FRGC31.7513.3810.05

21. 21 Security Analysis The reference binary templates are randomly generated, provide k bits entropy. Projection matrix M is unprotected. However, since M is only related to w Ω with equation and e is kept secret to attacker, M will not release useful information.

22. 22 Conclusions This paper has proposed a new method to generate a binary face template from a real valued face template. The discriminability of the extracted binary templates is optimized. The experimental results show that the proposed method has good performance. The security of the proposed algorithm is just the length k of the extracted binary template, which is quite sufficient when k is large.

23. 23 Q & A Thanks!

24. 24 Appendix