C - MMSP 2014

advertisement
MMSP
2014
4th Thematic Meeting
on Multimedia Signal
Processing
Forni di Sopra, Italy
February 17, 2014
A communication paradigm for
biometrics security and privacy
Emanuele Maiorana, Patrizio Campisi
emanuele.maiorana@uniroma3.it
Department of Engineering, University Roma Tre
Via Volterra 62, 00146, Rome, Italy
Biometric Recognition Systems
MMSP
2014
4th Thematic Meeting
on Multimedia Signal
Processing
Forni di Sopra, Italy
February 17, 2014
●
●
●
●
●
●
●
●
●
●
●
●
●
• Automatic systems performing people recognition
using their physical or behavioral characteristics
– applications
 law enforcement
o criminal identification
 access control
o physical or logical
– advantages
 cannot be lost or stolen
 improved security
– desired properties
 performance
 circumvention
 acceptability
E. Maiorana, P. Campisi - A communication paradigm for biometrics security and privacy
Security and Privacy Issues
• Main concerns affecting users’ acceptability
MMSP
2014
4th Thematic Meeting
on Multimedia Signal
Processing
Forni di Sopra, Italy
February 17, 2014
●
●
●
●
●
●
●
●
●
●
●
●
●
– security
 if compromised, biometrics cannot be replaced
– privacy
 misuse of data (function creep)
• Need for securing the employed biometric templates
– cryptography
 templates vulnerable during authentication
– hashing
 not error-tolerant, biometrics can vary
– template protection schemes
 templates modified for
o providing protection
o providing renewability
o guaranteeing performance
E. Maiorana, P. Campisi - A communication paradigm for biometrics security and privacy
Fuzzy Commitment
x
MMSP
2014
4th Thematic Meeting
on Multimedia Signal
Processing
Binarization
b
c
Encoding
m
Enrollment
Authentication
~
x
Hash
~
Binarization
b
v
~
c
Decoding
Hash
Decision
Forni di Sopra, Italy
February 17, 2014
●
●
●
●
●
●
●
●
●
●
●
●
●
• Most employed biometric protection scheme
– binarization of 𝐱  binary template 𝐛
– error correcting code 𝐜 XORed with 𝐛 to manage variability
– storage of helper data 𝐯
• Issues
– binding process biometrics/keys limited to the XOR
– binary block codes have low error correction capability (ECC)
 high False Rejection Rates (FRRs) due to biometrics variability
 low security, limited by the lenght k of the binary key 𝐦
A. Juels, M. Wattenberg, "A fuzzy commitment scheme", ACM CCS, 1999
E. Maiorana, P. Campisi - A communication paradigm for biometrics security and privacy
Code-Offset Sketch
v
MMSP
2014
4th Thematic Meeting
on Multimedia Signal
Processing
x
code C
~
x-x
Forni di Sopra, Italy
February 17, 2014
●
●
●
●
●
●
●
●
●
●
●
●
●
-x~
c
c~
• General construction
– c: codeword selected from a code C
– x: biometric template
– v = f(x,c): code-offset
 example: v = f(x,c) = x + c
 should not reveal information on neither x nor c
– 𝐜 = g(𝐱,v): revert binding operation
 user recognized iff 𝐱 ≃ 𝐱  𝐜 ≃ 𝐜
E. Maiorana, P. Campisi - A communication paradigm for biometrics security and privacy
Modulation Similarity
MMSP
2014
4th Thematic Meeting
on Multimedia Signal
Processing
c
Forni di Sopra, Italy
February 17, 2014
●
●
●
●
●
●
●
●
●
●
●
●
●
n
c~
• Modulation
– c: symbol selected from a constellation (ex.: QAM, PSK)
– n: noise added by the channel
 𝐧 = 𝐱 − 𝐱 in the code-offset scheme
– 𝐜: received corrupted symbol to be demodulated
• Proposed approach
– use of a modulation-like scheme for FC generalization
 improvement in both verification rates and security
E. Maiorana, P. Campisi - A communication paradigm for biometrics security and privacy
Proposed Scheme
x
MMSP
2014
4th
Thematic Meeting
on Multimedia Signal
Processing
Forni di Sopra, Italy
Hash
m
m
Turbo
Encoding
Modulation
c
f(x,c)
Enrollment
Authentication
Decision
Hash
~
m
February 17, 2014
●
●
●
●
●
●
●
●
●
●
●
●
●
Message
Generator
Channel
Joint Turbo Decoding and
Demodulation
~
v
g(𝐱,v)
c
x~
• Characteristics
– use of turbo-codes for managing intra-class variability
– codes modulated into s symbols of an L-points constellation
– binding expressed through a generic function f ()
 values in x may belong to an alphabet different than c’s one
 added noise characterized as n = c − c= g(x, f(x, c)) − f(x, c)
– joint demodulation and decoding of 𝐜 = g(𝐱,v)
 use of turbo-codes in soft-decoding modality
 ECC improved without user-specific information (privacy leakage)
E. Maiorana, P. Campisi - A communication paradigm for biometrics security and privacy
Practical Implementation
• Constellation
MMSP
2014
4th Thematic Meeting
on Multimedia Signal
Processing
Forni di Sopra, Italy
February 17, 2014
●
●
●
●
●
●
●
●
●
●
●
●
●
– Phase Shift Keying (PSK) modulation
 symbols in c as L possible points in a complex circle
• Binding function
 hard to say which symbols may generate a given value in v
o quantization of each element of
x in D possible values (D
> L)
o linear mapping of the values to the interval [- ; ]
o binding
v = f(x,c) = c · eix
𝐜 = g(𝐱, v) = v · e-i
̃x
c~
c
v
E. Maiorana, P. Campisi - A communication paradigm for biometrics security and privacy
Security evaluation
MMSP
2014
4th Thematic Meeting
on Multimedia Signal
Processing
• Measured through the entropy H(x/v)
– knowing v, having m directly provides x: H(m/v) = H(x/v)
 using a (n, k) turbo code, if v is known 𝐱 can be recovered from
𝑘
𝑛
only 𝑧 = log 𝐿 symbols (instead of 𝑠 = log 𝐿)
2
2
 𝐻 𝐱 𝑧 |𝐯 estimated with second-order dependency approximation
Forni di Sopra, Italy
February 17, 2014
●
●
●
●
●
●
●
●
●
●
●
●
●
𝑃 𝐱𝑧 =
𝑧
𝑖=1 𝑃
𝐱 𝑢𝑖 𝑧 |𝐱 𝑡(𝑢𝑖 ) 𝑧 , 1𝑡 𝑢𝑖 <𝑢𝑖 , 𝐮 = 𝑢𝑖 , 1  𝑖  z
𝑧
𝐻 𝐱 𝑧 |𝐯 = min min
𝑧∈𝑍
𝐮
𝐻 𝐱 𝑢𝑖 𝑧 , 𝐱 𝑡(𝑢𝑖) 𝑧 |𝐯 − 𝐻 𝐱 𝑡(𝑢𝑖 ) 𝑧 |𝐯
𝑖=1
– ideal biometric representations 𝐱 should possess
 low intra-class variability, for managing it with feasible ECCs
 a large number of features, for using codes with large k
 independent features, for the 𝑧
most-correlated ones
determine the security
E. Maiorana, P. Campisi - A communication paradigm for biometrics security and privacy
Application: On-line Signatures
• Biometric representation
MMSP
2014
4th Thematic Meeting
on Multimedia Signal
Processing
– use of Universal Background Models (UBMs)
 users’ templates obtained adapting a person-independent model
o use of Hidden Markov Models (HMMs) as global model

adaptation of the Gaussians mean values with a user-independent matrix P
mu = mUBM + P·xu
Forni di Sopra, Italy
February 17, 2014
●
●
●
●
●
●
●
●
●
●
●
●
●
– characteristics
 theoretically uncorrelated features
 possible large feature number: 4800
o significant intra-class variability

s most stable features selected in a training phase
• no information-leakage
• Database
– samples taken from the MCYT on-line signature DB
 25 genuine and 25 skilled forgeries for each of 100 users
E. Maiorana, P. Campisi - A communication paradigm for biometrics security and privacy
Experimental results
• Fuzzy commitment
MMSP
2014
– BCH codes  unacceptable performance
s
FRR
FAR
4th Thematic Meeting
on Multimedia Signal
Processing
Forni di Sopra, Italy
February 17, 2014
●
●
●
●
●
●
●
●
●
●
●
●
●
511
50.0
0.3
8.7
H(xZ|v)
1023
61.5
0.2
9.6
2047
77.8
0.0
10.4
4095
93.6
0.0
11.3
– turbo codes on binary data (L=D=2)  still high FRR

no soft-decoding
s
511
4.6
8.5
22.9
FRR
FAR
H(xZ|v)
1023
7.4
3.9
50.7
2047
15.0
1.6
108.1
4095
32.4
0.4
221.5
• Proposed approach
s
L
D=4
D=8
2047
FRR
FAR
H(xZ|v)
FRR
FAR
H(xZ|v)
2
2.1
22.3
79.8
1.3
67.2
61.0
4
6.1
8.7
184.1
0.5
74.5
126.2
4095
2
6.0
7.8
166.7
4.1
44.2
127.9
4
13.2
2.2
378.1
1.8
54.3
274.2
– increasing L improves FAR and security, worsening FRR
– increasing D improves FRR, worsening FAR and security
E. Maiorana, P. Campisi - A communication paradigm for biometrics security and privacy
Application: Iris
• Biometric representation
MMSP
2014
– use of Daugman’s rubber sheet model
 iris segmented and normalized to create a rectangular template
o phase information of the Gabor filtering retained as features
4th Thematic Meeting
on Multimedia Signal
Processing
Forni di Sopra, Italy
February 17, 2014
●
●
●
●
●
●
●
●
●
●
●
●
●
– characteristics
 highly correlated features (severe issue for security)
 possible large feature number: I x J = 240 x 20 = 4800 features
 significant intra-class variability
o intra-class variability reduction with a user-independent mask

keeps regions where occlusions (eyelids) are not encountered (s = 2048)
• Database
– samples taken from the CASIA-v4 iris DB
 2251 images taken from 395 irises, from 249 subjects
E. Maiorana, P. Campisi - A communication paradigm for biometrics security and privacy
Experimental results
• Fuzzy commitment vs Proposed approach
MMSP
2014
4th Thematic Meeting
on Multimedia Signal
Processing
BCH Codes
Fuzzy
Commitment
Turbo Codes
Forni di Sopra, Italy
February 17, 2014
●
●
●
●
●
●
●
●
●
●
●
●
●
Proposed
Approach
L=2
L=4
D=2
D=4
D=8
D=2
D=4
D=8
FRR
48.8
53.7
60.4
71.3
13.7
13.8
13.9
14.3
FAR
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
13
31
47
71
268
310
368
450
H(xZ|v)
5.7
13.9
20.7
31.2
117.7
140.4
163.5
200.4
13.4
4.84
3.00
13.6
8.92
0.0
0.1
0.3
0.0
0.0
132
132
132
268
268
67.4
18.6
11.6
117.7
46.7
k
• Recognition accuracy comparison
E. Maiorana, P. Campisi - A communication paradigm for biometrics security and privacy
Conclusions
MMSP
2014
4th Thematic Meeting
on Multimedia Signal
Processing
• Proposal of a template protection scheme inspired by
the digital communication paradigm
– modulation constellations for symbol representation
– turbo-decoding employed in soft-modality
 generalization of the fuzzy commitment scheme
Forni di Sopra, Italy
February 17, 2014
●
●
●
●
●
●
●
●
●
●
●
●
●
• Tests with on-line signatures and iris biometrics
– great flexibility in selecting the operating conditions
– improved performance in verification rates and security
 achieving low FRR and high security are conflicting requisites
o proper selection of parameters L and D
• Future developments
– applications to other biometrics
– analysis of multi-biometrics systems
E. Maiorana, P. Campisi - A communication paradigm for biometrics security and privacy
Bibliography
MMSP
2014
4th Thematic Meeting
on Multimedia Signal
Processing
Forni di Sopra, Italy
• E. Maiorana, P. Campisi, “Fuzzy Commitment for Function based
Signature Template Protection”, IEEE Signal Processing Letters,
Vol. 17, No. 3, pp. 249 - 252, March 2010
• E. Maiorana, “Biometric cryptosystem using function based online signature recognition”, Expert Systems with Application,
Elsevier, Vol. 37, No. 5, pp. 3676 - 3684, May 2010
February 17, 2014
• E. Argones, E. Maiorana, J. A. Castro, and P. Campisi, “Biometric
template protection using universal background models: An
application to online signature,” IEEE Transactions on Information
Forensics and Security, 7(1), pp. 269–282, January 2012
• E. Maiorana, D. Blasi, P. Campisi, “Biometric Template Protection
using Turbo Codes and Modulation Constellations,” IEEE WIFS,
December 2012
• E. Maiorana, P. Campisi, A. Neri, “Iris Template Protection Using
A Digital Modulation Paradigm”, IEEE ICASSP, May 2014
E. Maiorana, P. Campisi - A communication paradigm for biometrics security and privacy
MMSP
2014
4th Thematic Meeting
on Multimedia Signal
Processing
Forni di Sopra, Italy
Thanks for your attention!
February 17, 2014
Questions?
emanuele.maiorana@uniroma3.it
E. Maiorana, P. Campisi - A communication paradigm for biometrics security and privacy
Download