Privacy-Preserving Face Recognition

Zekeriya Erkin1, Martin Franz2, Jorge Guajardo3,

Stefan Katzenbeisser2, Inald Lagendijk1, and Tomas Toft4

Introduction

Alice Bob

Owns a face imageIs neither willing to share the image nor the detection result

Owns a face databaseIs not willing to reveal his data

Run face recognition to determine

whether the face image is in database？

FIND

Paillier cryptosystem

• additively homomorphic public-key encryption schemes

• [a + b] = [a][b],• [ab] = [a]b. ( b is a constant)

Face Recognition

• Run Principal Component Analysis (PCA) from a set of criminal images to obtain eigenface

• Projects face images onto eigenfaces.

PrincipalComponent Analysis (PCA)

• Θ1,Θ2, . . . , ΘM : vectors of length N• average of the training images• covariance matrix• Run PCA• To determine the face space, we selectK << M eigenvectors u1, . . . , uK that correspond to the K largest eigenvalues.

criminal projection

• Θ1,Θ2, . . . , ΘM are projected onto the subspace spanned by the basis u1, . . . , uK to obtain their feature vector representation Ω1, . . . , ΩM.

Suspect Projection

• input image Γ

Progection in the encrypted domain

• input image

Calculating distances

Client’s

Calculating distances(conti)

BobAlice

Calculating distances

•DEMO~

Paillier cryptosystem

• Two large prime number p, q• n = p*q• Select random integer g • •

• Encryption• Decryption

Private key

Public key

Paillier cryptosystem

• Homomorphic addition of plaintexts

• Homomorphic multiplication of plaintexts