FACE FLASHINGwp.internetsociety.org/ndss/wp-content/uploads/sites/25/...FACE FLASHING: A SECURE...

Post on 08-Jul-2018

219 views 0 download

transcript

FACE FLASHING: A SECURE LIVENESS DETECTION PROTOCOL BASED ON LIGHT REFLECTIONS

Di Tang1, Zhe Zhou2, Yinqian Zhang3, Kehuan Zhang1

The Chinese University of Hong Kong1

Fudan University2

The Ohio State University3

Face-based Authentication Will Become Popular

2

Online payment

Door entrance

ATM withdraw

Phone unlock

� Easy-obtained faces

3

Face Recognition Is Not Enough

4

�  Easy-obtained faces

�  High-resolution printers/screens

�  Powerful CPUs/GPUs

�  Developed technologies

5

Face Recognition Is Not Enough

6

Liveness Detection Is Necessary

Detect whether the subject under authentication is a real human

7

Liveness Detection Is Hard to Be Done Right

Texture extraction methods:

- Local Binary Pattern (LBP) -  2D Fourier Spectra

-  …

High-resolution screen will fail it. ---- It can outputs any patterns you want

8

Challenge-response protocols:

-  Eye blink

-  Expression

-  Head movement

- Speaking

Liveness Detection Is Hard to Be Done Right

9

Human Reaction Time

Machines can do109 flops, in 260MS

www.humanbenchmark.com

10

Machines Are Powerful

3D reconstruction

Face morphing

11

Machines Are Powerful

Expression synthesizing

12

Fundamental Problem ?

13

Fundamental Problem ?

No strong security guarantee!

Details

Trembling Ability

Precision

14

Weakness of Human Reactions

Limited speed Uncertainty Smart device + Screen can fail it

2D dynamic attacks (e.g., Media-based Facial Forgery)

15

What We Want to Do?

Solid stone to build a secure protocol

Human reaction

Relieve threats from 2D dynamic attacks

16

Light reflection

Non-digital physical

17

Features of Light Reflection

Fastest in the universe -- No computers can generate fake responses at the same speed, no matter how powerful it will be Without human reaction Can capture rich information -- 3D shape -> eyes, nose -- Texture -> skin vs. non-skin

E: Illumination R: Reflectance S: Sensor response function λ : Wave length x : position of a given point

We will separately consider R,G,B channels. There are no inter-effect among them, if we use the raw data (before AWB).

Reflection Model

18

19

E: Incoming light R: Reflectance

Get reflectance: Get illumination:

The reflections is determined by incoming light Without knowing the incoming light, it is impossible to pre-calculate the reflected light.

Reflection Model

To check face To check time

20

Things to verify: 1. Response time 2. Face information 3. Expressions

Design

21

� Challenging!!

� Reflections happen at speed of light

� But camera is not

� Limited by the refreshing speed

� à around 30 fps

� Does it mean powerful attackers with high speed camera and displaying devices can bypass?

Verifying the Timing is Difficult

22

WORKING DETAILS OF CAMERA

Working Details of Camera

23

WORKING DETAILS OF CAMERA

Working Details of Camera

column column column

column

Col 0

Col 1

Col 2

Col 3

Col N-1

Col N

Anytime, there are always sensors awake

24

Detecting tiny differences in time is possible

25

Both camera and LCD monitor work in a scanning pattern. So what will happen?

Working Details of Screen

Assumption: No modification can be added to the buffer that is being displayed

26

Partially Captured Images

Camera Screen

27

How to verify?

Lighting challenge Background challenge

28

Challenges

29

Response and Challenge

Get challenges:

Lighting area Lighting area

Mirror

30

Calculate the Location

The Challenge image (with lighting area)

Corresponding region

Camera

Forgery -> Delay -> Wrong location

Accumulation:

31

Get reflectance:

Put it into a Neural network for classification

Face Feature Verification

Evaluation

33

SENSITIVITY TO FAKE RESPONSES

Sensitivity to Forged Responses

34

Timing: Camera VS. Mirror

Mirror’s

Laptop’s

35

Face Feature

36

Robustness

Our method will force adversaries to use “3D Dynamic Attack” which is more expensive

Our method could not handle 3D dynamic attack

twins, silicone masks

31

Discussion

Our implementation just used 8 different colors Our implementation needs several seconds to accomplish once authentication Using ‘albedo curve’ may handle 3D dynamic attacks Combine with face recognition algorithm could enhance efficiency and effectiveness

32

Discussion

Face Flashing protocol Effective and efficient method on timing and face verifications

Prototype and empirical evaluations 33

Summary

Q & A

Thanks