1. Anonymous Biometrics: Privacy Protection of Biometric Templates Pim Tuyls, E. Verbitskiy, D. Denteneer, J.P. Linnartz, J. Goseling, T. Ignatenko Pim.Tuyls@philips.com Philips Research Eindhoven The Netherlands

2. 2 Overview Introduction Challenge Literature and Related Topic Information-Theoretic model Secrecy Extractor Requirements Bounds Examples “General” Theory Experiments Summary

3. 3 Introduction Biometric Identification (fingerprints, iris, speech) - is often used to identify people - is often part of a security system - uses databases containing Ref. Information (Templates) Advantages Convenience can not be lost or forgotten easy to use Uniqueness unique for a human being Offers therefore a very attractive alternative to e.g. passwords

4. 4 Risks - Forgeability - Impersonation by Artificial Biometrics - Once Compromised Compromised Forever -Theft of Identity (Stolen Biometrics) - Sensitive Information - Fingerprints contain Genetic Information - Retina reveals susceptibility for Strokes and Diabetes Additional Problem - Noisy: Biometric data are obtained through noisy measurements PRIVACY

5. 5 ARCHITECTURE ASSUMPTIONS Database public Channel public Sensor trusted ATTACKS - Outside (on database) - Eavesdropping of Communications - Inside (on database): Malicious owner (Verifier) - Fingerprints left on glasses, door handles (not discussed today) Database Sensor Template Channel

6. 6 Possible Constructions: - Encryption (implies a decryption key at verifier site) - One-Way Function Idea: Build a scheme similar to the one used for password protection Solution Secure Storage of Biometric Templates, Against Outside and Inside Attacks Secure Communication over the Channel (prevent eavesdropping)

7. 7 CHALLENGE: Integration of Cryptographic Techniques with Noisy Inputs One-Way Functions are very sensitive to small changes in the input data database matching F F

8. 8 Literature - Schneier - Davida, Frankel and Matt, (Private biometrics) - Juels and Wattenberg (Fuzzy Commitment) - Ratha, Connell, Bolle (Cancelable Biometrics) - Juels, Sudan (fuzzy vault) - Linnartz, Tuyls (Shielding functions, AVBPA 2003) - Verbitskiy, Tuyls, Denteneer and Linnartz (Benelux 2003) - Goseling, Tuyls submitted to ISIT2004 Related Topic - Biometric Key Generation (Soutar)

9. 9 Information Theoretic Model Biometrics X n are modeled as random variables with distribution(enrollment) Authentication measurements Y n, modeled as observations through a noisy channel

10. 10 Generate Common Secret S from X n and Y n (Common Randomness) Helper data W Secrecy Extractor Database: ID, W, F(S) matching F F G G ’)? EXACT MATCH: F(S)=F(S’)? Enrollment Authentication F(S)

11. 11 Terminology A functionis called a  -contracting function: if for all X there exist a W s.t probabilistic norm Versatile function: for all S  0,1  k and all X  R n, there exists a vector W  R m such that:  -Revealing function: 

12. 12 Requirements A reliable biometric authentication system that protects privacy has to satisfy the following requirements:  -contracting Versatile  -revealing: Correctness: Protection against a dishonest verifier who has Access to the database (compare with passwords)

13. 13 Implications Proposition 1: If W is constant, i.e. G(Y,W)=C(Y) then either  =0, or G(Y,W) is a constant independent of Y. Corollary: In order to have a robust, versatile function G=G(X,W), W must depend on X

14. 14 Implications Proposition 2 : Let S be a binary string derived from X and Y by communicating helper data W as described in the protocol: Extends also to the continuous case! (Approximation argument)

15. 15 EXAMPLES Three kinds of proposed schemes: Based on Quantized Index Modulation Error Correcting Code-scheme Significant Components

16. 16 Example: Significant Components Assumption: Orthogonal Transformation (Fisher, PCA): Define: where  i are orthonormal vectors Theorem (Fisher, PCA): The  i can be constructed such that they are independent, normally distributed random variables with zero mean

17. 17 The Scheme I: Robustness Idea: Select  -components with large absolute values to guarantee robustness to noise Choose a small positive number  and define Theorem: Let  be the fraction of average number of large comps then, if there is a sufficient amount of energy in the system,  is “large”, moreover

18. 18 The Scheme II: Versatility Versatility: Given s i, search for index i j such that: (feasibility) The set of feasible secrets: Theorem: If k=  1 n with  1 =  /10, then with large probability is a large set

19. 19 The Scheme III: Helper Data Given a secret S=(s 1,…,s k ) the helper data W is determined. W picks up the correct components of X in  -basis Helper data: W(X) is a k  n matrix, its j-th row is given by  -contracting function:

20. 20 Information Revealing Theorem: The proposed scheme is zero-revealing: Moreover,

21. 21 General Construction SEC: Tuple of encoding regions (SEC: Secure Extraction Code) such that, is the collection of SECs s.t.

22. 22 Secure Biometric Authentication Scheme (SBA) 1.Enrollment measurement X n 2.Select a code in W indicates the selected code 3.The Secret S is index of that coding region where X n belongs to 4. A One-Way Function F is applied to S. 5.W and F(S) are stored in the database together with the Id. ENCDEC 1 2 3

23. 23 Authentication: 1. An individual makes an Id claim 2. W and is sent to the decoder 3. The SEC C(W) is used to derive the secret as follows, 4. 5. F(S’) is computed 6. Check: F(S’)=F(S) This construction achieves the earlier mentioned capacities at the same time (Asymptotically)!

24. 24 Experiments - Biometric: Measuring the headphone-to-ear-canal-Transfer Functions - First dataset: 45 Individuals, 8 Measurements per person - Second dataset: 65 Individuals, 8 Measurements per person - 6 Measurements for training, 2 for authentication - Tested scheme: significant components - FRR decreases as  increases - FAR decreases as secret length increases - Secret length decreases as  increases

25. 25 “Ear canal” Biometrics = Headphone-to-Ear Transfer Function White noiseError H(z) W(z) +

26. 26 Headphone-to-Ear Transfer Function: 1 ear, population (45x8)

27. 27 Results: Principal Component Transform First dataset

28. 28 Second dataset Combination of schemes

29. 29 Summary We have described a general set-up and examples for biometric authentication/key generation schemes that satisfy the following properties: - Robust to noise - Versatile - Zero-revealing - Privacy protection