HEFT: Homomorphically Encrypted Fusion of Biometric Templates
Luke Sperling† \dagger † , N. Ratha‡ \ddagger ‡ , A. Ross† \dagger † , V. Boddeti† \dagger †
† \dagger † Michigan State University, ‡ \ddagger ‡ University at Buffalo
12th October, 2022
IJCB 2022
Fusion of Biometric Information
"A comprehensive overview of biometric fusion."Information Fusion, 2019"
"Deep learning approach for multimodal biometric recognition system based on fusion of iris, face, and finger vein traits." Sensors, 2020
Information Leakage from Representations
Attacks on Face templates
"Assessing Privacy Risks from Feature Vector Reconstruction Attacks," arXiv:2202.05760
Face reconstruction from template
"On the reconstruction of face images from deep face templates," TPAMI, 2018
Finger vein reconstruction from binary templates
"Inverse Biometrics: Reconstructing Grayscale Finger Vein Images from Binary Features," IJCB, 2020
Encryption: The Holy Grail?
Data encryption is an attractive option
protects user's privacy
enables free and open sharing
mitigate legal and ethical issues
Encryption scheme needs to allow computations directly on the
encrypted data.
Solution: Homomorphic Encryption
Key Idea of Homomorphic Encryption
Ring Learning with Errors
op
plaintext
ciphertext
x x x
$(x + e_1) \mbox{ mod } t$
y y y
$(y + e_2) \mbox{ mod } t$
+ + +
x + y x+y x + y
$(x+y + e_3') \mbox{ mod } t$
× \times ×
x × y x\times y x × y
$(x\times y + e_4'') \mbox{ mod } t$
FHE in Biometrics
"Secure Face Matching Using Fully Homomorphic Encryption,", BTAS 2018
"HERS: Homomorphically Encrypted Representation Search,", TBIOM 2022
Focussed on protecting database of templates.
Allows match score computation in the encrypted domain.
HEFT
Homomorphically Encrypted Fusion of Biometric Templates
HEFT: Overview
HEFT: Concatenation
Homomorphic Concatenation
HEFT: Linear Projection
Linear Projection
Naive
Hybrid
SIMD
Linear Projection Comparison
Computational Complexity
Space Complexity
Hybrid
Pros: Low memory and runtime overhead
Cons: Scales linearly with number of samples
SIMD
Pros: Scales well with number of samples
Cons: High memory and runtime overhead
HEFT: Feature Normalization
ℓ 2 \ell_2 ℓ 2 -Normalization of Vector
u ^ = u ∥ u ∥ 2 → \hat{\mathbf{u}} = \frac{\mathbf{u}}{\|\mathbf{u}\|_2} \quad \rightarrow \quad u ^ = ∥ u ∥ 2 u → division† \dagger †
where
∥ u ∥ 2 = ∑ i = 1 d u i 2 → \|\mathbf{u}\|_2 = \sqrt{\sum_{i=1}^d u_i^2} \quad \rightarrow \quad ∥ u ∥ 2 = ∑ i = 1 d u i 2 → square-root† \dagger †
† \dagger † : problematic operations for FHE
Inverse Square Root: Polynomial Approximation
1 x = ∑ i = 1 6 a i x i \frac{1}{\sqrt{x}} = \sum_{i=1}^6 a_i x^i x 1 = i = 1 ∑ 6 a i x i
FHE-Aware Learning
Account for the limitations of FHE to improve performance
FHE is limited to specific operations on encrypted data.
Normalization is not directly computable - need to approximate .
Approximation is a source of error and hence a loss of matching performance
We incorporate approximate normalization into our training of the projection matrix to recover performance
Loss Function
Main Idea: FHE-Aware Learning
L o s s = λ ∑ M d ( c i , c j ) ∣ M ∣ ⏟ P u l l + ( 1 − λ ) ∑ V [ m + d ( c i , c j ) − d ( c i , c k ) ] + ∣ V ∣ ⏟ P u s h Loss = \lambda \underbrace{\frac{\sum_M d(\mathbf{c}_i, \mathbf{c}_j)}{|M|}}_{ \color{orange}{Pull} } + (1-\lambda)\underbrace{\frac{\sum_{V}[m + d(\mathbf{c}_i, \mathbf{c}_j) - d(\mathbf{c}_i, \mathbf{c}_k)]_{+}}{|V|}}_{ \color{orange}{Push} } L oss = λ P u ll ∣ M ∣ ∑ M d ( c i , c j ) + ( 1 − λ ) P u s h ∣ V ∣ ∑ V [ m + d ( c i , c j ) − d ( c i , c k ) ] +
where d ( c i , c j ) = 1 − P f ( c i ) ⏟ a p p r o x i m a t i o n ⋅ P f ( c j ) ⏟ a p p r o x i m a t i o n d(\mathbf{c}_i, \mathbf{c}_j) = 1-P\underbrace{f(\mathbf{c}_i)}_{ \color{cyan}{approximation} } \cdot P\underbrace{f(\mathbf{c}_j)}_{ \color{cyan}{approximation} } d ( c i , c j ) = 1 − P a pp ro x ima t i o n f ( c i ) ⋅ P a pp ro x ima t i o n f ( c j )
f ( ⋅ ) f(\cdot) f ( ⋅ ) approximates the inverse norm of a vector.
Experimental Setup
Cross-Posed Labelled Faces in the Wild
Down
Dog
Bird
Backward
Google Speech Commands
Synthetic fusion dataset by randomly pairing classes.
10,760 samples over 188 classes.
Fusion Improves Performance, Reduces Dimensionality
0.8401 0.855 0.9253 0.9508 CPLFW (Dataset 1) GSC (Dataset 2) Concatenation HEFT 0 0.2 0.4 0.6 0.8 1
AUROC
Fusion improves performance:
Face by 11.07%
Voice by 9.58%
Dimensionality Reduction: 512 D → 32 D 512D \rightarrow 32D 512 D → 32 D (16× \times × compression)
Comparison of Normalization Methods
Computational Complexity
Concatenation Projection Normalization Preprocessing 0 200 400 600 800 1000
Poly (deg=2) Poly (deg=6) Goldschmidt Enrollment
Projection is costliest operation
Concatenation Projection Normalization Matching 0 2000 4000 6000 8000
Poly (deg=2) Poly (deg=6) Goldschmidt Authentication
Projection is costliest operation
Summary
Introduces the first multimodal feature-level fusion system in the encrypted domain.
Improves performance of original templates while reducing their dimensionality.
Incorporates polynomial approximation for approximate normalization.
Incorporates FHE-Aware Learning to improve performance.
HEFT: Encrypted Biometric Fusion
Resume presentation
HEFT: Homomorphically Encrypted Fusion of Biometric Templates Luke Sperling $\dagger$ , N. Ratha $\ddagger$ , A. Ross $\dagger$ , V. Boddeti $\dagger$ $\dagger$ Michigan State University, $\ddagger$ University at Buffalo 12th October, 2022 IJCB 2022