Face anti-spoofing using Image Quality Assessment

Size: px

Start display at page:

Download "Face anti-spoofing using Image Quality Assessment"

Phillip Thomas
6 years ago
Views:

1 Face anti-spoofing using Image Quality Assessment

2 Speakers Prisme Polytech Orléans Aladine Chetouani R&D Trusted Services Emna Fourati

3 Outline Face spoofing attacks Image Quality Assessment Proposed method q General framework q Feature extraction q Experimental results

4 Outline Face spoofing attacks Image Quality Assessment Proposed method q General framework q Feature extraction q Experimental results

Face spoofing attacks Goal : evaluate face recognition

Maximum level of user control and privacy: Complied with

and match on user device No database, no server Selected

5 Face spoofing attacks Goal : evaluate face recognition technologies Targeted implementation: Mobile devices Maximum level of user control and privacy: Complied with GDPR (General Data Protection Regulation) Record, storage and match on user device No database, no server Selected SDK Hybrid method Picture and face features Limits Spoofing is far too simple

6 Face spoofing attacks Printed photo/digital photo/video 3D mask Easy to reproduce 2 photos (frontal and profile) Online: Affordable (299$)

7 Face spoofing attacks State-of-the-art solutions Liveness detection techniques (eye blinking, facial expression changes, etc.) Motion analysis techniques Facial appearance analysis techniques Contextual information techniques User-interaction based methods (eg. Turn head left/right, blink, smile, etc.) Multimodal biometric approaches Our works Multimodal biometric authentication Combining voice recognition and lip movement Combining PIN, face and BioTyping Liveness detection based on Image quality assessment

8 Face spoofing attacks State-of-the-art solutions Liveness detection techniques (eye blinking, facial expression changes, etc.) Motion analysis techniques Facial appearance analysis techniques Contextual information techniques User-interaction based methods (eg. Turn head left/right, blink, smile, etc.) Multimodal biometric approaches Our works Multimodal biometric authentication Combining voice recognition and lip movement Combining PIN, face and BioTyping Liveness detection based on Image quality assessment

9 Outline Face spoofing attacks Image Quality Assessment Proposed method q General framework q Feature extraction q Experimental results

10 Image Quality Assessment Image Quality Measures (IQMs) Full-Reference IQA Error Sensitivity: Specific features (Pixel Differences, Correlation ) Structural Similarity: Inspired by the Human Visual System (HVS) Information Theoretic: Mutual Information/ Histogram e.g: MSE, PSNR à Low complexity e.g: SSIM à Most widely used e.g: VIF à Heavy computation (using simplified versions)

11 Image Quality Assessment Image Quality Measures (IQMs) No-Reference/ Blind IQA Distortion specific: Assumptions on the distortion type(s) Training-based Specific pre-trained features Natural Scene Statistics Deviation from the corpus of Natural undistorted images e.g: HLFI à Blur and random noise e.g: GM-LOG-BIQA à Gradient-Magnitude & Laplacian-of-Gaussian e.g: BIQI à Distorted Image Statistics

12 Proposed method Calculation approach Image Quality Measures (IQMs) Reduced-Reference IQA Using Only partial information from the reference image e.g: RRED à Differences between the entropies of wavelet coefficients

13 Outline Face spoofing attacks Image Quality Assessment Proposed method q General framework q Feature extraction q Experimental results

14 Proposed method General framework General diagram Training set IQA measures Real Fake Error rates

15 Proposed method General framework Frame extraction Frame 1 Face/Background motion (FBM) Frame 2 FBM= Motion(Face,F1,F2,th)/ Motion (Background, F1,F2,th) Motion(RoI,F1,F2,th)= x,y δ(d(x,y) th) /SD High FBM Extract frame Low FBM Move to next frame - th: Pixel-wise difference threshold - δ: Dirac delta function - F1, F2: Two consecutive frames - RoI: Face / Background - SD: Pixel number of the considered RoI

Proposed method General framework Frame extraction Frame 1 Face/Background motion (FBM) Frame 2 FBM= Motion(Face,F1,F2,th)/ Motion (Background, F1,F2,th) Motion(RoI,F1,F2,th)= x,y δ(d(x,y) th) /SD J

16 Proposed method General framework Frame extraction Frame 1 Face/Background motion (FBM) Frame 2 FBM= Motion(Face,F1,F2,th)/ Motion (Background, F1,F2,th) Motion(RoI,F1,F2,th)= x,y δ(d(x,y) th) /SD J Relevant differences between the frames J Liveness-related motion cues (face + throat) J Computational efficiency - th: Pixel-wise difference threshold - δ: Dirac delta function - F1, F2: Two consecutive frames - RoI: Face / Background - SD: Pixel number of the considered RoI

17 Proposed method General framework Training set IQA measures Real Fake Error rates

18 Outline Face spoofing attacks Image Quality Assessment Proposed method q General framework q Feature extraction q Experimental results

19 Proposed method Feature extraction IDIAP Research Institute, 2014 (1) - 25 IQMs(21 FR & 4 NR) à LDA classification: HTER 15.2% IDIAP & GRADIENT Research center, 2016: (2) - A subset of 18 IQMs (17 FR and 1 NR) à LDA classification : HTER 9.78% Our work - Simulating other promising IQA metrics - Combining the best ones with the 18-feature selection (1)Galbally,J.; Marcel, S.; Fierrez, J.: Image quality assessment for fake biometric detection: Application to iris, fingerprint and face recognition, IEEE Trans. on Image Processing, vol. 23, no. 2, pp , 2014 (2) Costa-Pazo, A. et.al.: The replay-mobile face presentation-attack database, International Conference on Biometrics Special Interests Group (BioSIG), 2016.

20 Proposed method Feature extraction Choice of additional metrics: 3 additional NR IQMs Compromise between: Classification performance Gaussian derivative Distorted Image Statistics Computational cost (Mobile implementation) GMLOGBIQA (1) BRISQUE (2) BIQI (3) (1) Xue,W. et.al.: Blind Image Quality Prediction Using Joint Statistics of Gradient Magnitude and Laplacian Features, Trans. on Image Processing, IEEE selben Jahr. Format-Verlag, (2) Mittal, A.; Moorthy, A. K. ; Bovik,A. C.:"No Reference Image Quality Assessment in the Spatial Domain", IEEE Transactions on Image Processing, 2011 (3) Moorthy, A. K.; Bovik, A. C.: "A Modular Framework for Constructing Blind Universal Quality Indices", submitted to IEEE Signal Processing Letters, 2009.

21 Proposed method Feature extraction Training set IQA measures Real Fake Error rates

22 Proposed method Feature extraction Our model Frame 1 Frame 2 Frame 3 Frame 4 Frame N... FR IQA FR IQA FR IQA FR IQA NR IQA NR IQA NR IQA NR IQA Feature vectors for N-1 frames à For each video

23 Outline Face spoofing attacks Image Quality Assessment Proposed method q General framework q Feature extraction q Experimental results

24 Proposed method Experimental results Performance indicators: False Genuine Rate (FGR) False Fake Rate (FFR) Sample: Fake Prediction: Real Sample: Real Prediction: Fake Half Total Error Rate: HTER = (FGR +FFR)/2

25 Proposed method Experimental results Replay Attack (1): 1300 videos (320x240) 50 subjects Different scenarios Lamp light Real Printed Phone Tablet Day light (1) Chingovska I. et.al.: "On the Effectiveness of Local Binary Patterns in Face Antispoofing", IEEE BIOSIG 2012

Proposed method Experimental results Replay Mobile (2): 1190 videos (1280 x 720) 40 subjects Phone Screen, light on Screen, light off Print, light on Print, light off Different scenarios à

26 Proposed method Experimental results Replay Mobile (2): 1190 videos (1280 x 720) 40 subjects Phone Screen, light on Screen, light off Print, light on Print, light off Different scenarios à Well suited to Mobile acquisition technologies Tablet (2) Costa-Pazo, A. et.al.: The replay-mobile face presentation-attack database,int. Conf. on Biometrics Special Interests Group (BioSIG) 2016

27 Proposed method Experimental results LDA classification on Replay Attack IQMs Video frames FGR FFR HTER-f HTER-v Initial 18 IQMs All 2,21 16,67 9,44 6,25 Motion 3,15 16,40 9,78 7,91 Final 21 IQMs All 8 e -04 0,009 0,005 0 Motion 0,002 0,045 0,023 0

28 Proposed method Experimental results LDA classification on Replay Mobile IQMs Video frames FGR FFR HTER-f HTER-v Initial 18 IQMs All Motion Final 21 IQMs All Motion

29 Conclusion Face antispoofing Motionbased frame extraction FR-IQA measures between the frames Other robust NR metrics Future works Testing other databases Deep Learning

31 Appendix LDA classification on Replay Attack IQMs Video frames Scaling FGR FFR HTER-f HTER-v LDA SVM LDA SVM LDA SVM LDA SVM Initial 18 feature-set Our 21 feature-set All Motion All Motion without Z-score without Z-score without 8e Z-score without Z-score

32 Appendix LDA classification on Replay Mobile IQMs Video frames Scaling FGR FFR HTER-f HTER-v LDA SVM LDA SVM LDA SVM LDA SVM Initial 18 IQMs Final 21 IQMs All Motion All Motion without Z-score without Z-score without Z-score without Z-score

Biometric Security System for Fake Detection using image Quality Assessment Techniques

Biometric Security System for Fake Detection using image Quality Assessment Techniques Rinu Prakash T, PG Scholar Department of Electronics & Communication Engineering Ilahia College of Engineering & Technology