Correlation Based Fingerprint Liveness Detection

Correlation Based Fingerprint Liveness Detection

Under The Guidance of: Ashish PawarProf. Mrs. Vidya Patil 30349

Biometrics

Bio = “life” metrics = “measure” To analyse physiological and behavioural characteristics Biometric is based on anatomic uniqueness of a person Three ways authentication can be done:1. Something the person knows(password)2. Something the person has(key , special card)3. Something the person is(fingerprint , footprint)

Biometric Identifiers

Fingerprint Recognition Iris Recognition Face Detection Voice Recognition DNA

Why Fingerprint?

Unique Reliable Fast and easy capture

Liveness It is a dichotomy of feature space into living and non-livingThere are essentially three different ways to introduce liveness detection into biometric systems:1. Using extra hardware to acquire life signs2. Using the information already captured by the system to detect life

signs.3. Using liveness information inherent to the biometric.

Relevance of liveness Detection

Systems requiring high security level

Because fingerprints cannot be changed

Accurate authentication

Features and patterns

Features

1. Sweat Pores

2. Ridges and valleys

3. Perspiration Patterns

1. Arch

2. Loop

3. Whorl

Fingerprint Spoofing Techniques

Direct CastingDirect Finger is used as source. Indirect CastingFingerprint obtained from secondary sources.

Existing Methods

Local Binary Pattern(LBP) Pore Detection Power Spectrum Ridges Wavelet Valley Wavelet

Comparison between existing methods

The above table gives the FerrFake rate

Comparison between existing methods

The above table gives the FerrLive rate

Overview of the proposed method

Used for distinguishing spoofed and live. Makes use of correlation Based on the underlying assumption of same class Automatic Adaptation to new fingerprint images

Algorithms used

Partial Least Square Method1. It is mostly used for prediction2. Used when factors are more as compared to observed values3. In this method used for correlation

Working of PLS

X = TPT + E

Y = UQ T + F

Factors Responses

Population

Figure 3: Indirect modeling

T U

factors response

Support Vector Machine(SVM)

1. SVM is a supervised learning method that generates ip/op mapping function from a set of labelled training data

2. It basically calculates probability of input applied .

3. In this method it is combined along wih the classifier algorithms like GMM,GC and QDA

FrameWork Architecture

Working

Feature Extraction phase

Features are extracted using specific descriptor like LBP and LPQ. Correlation Phase

Correlation is performed using PLS

Here it is also used to model relation using below equation:Lf = TPT + E

Sf = UQ T + F

Where Lf anf Sf are two matrices after feature extraction

SVM PhaseProbablities are calculated for live and spoof classesDenoted by xl,xs. Generative Classifier PhaseHere using above probabilities result is generated

(1)

Comaprison with existing Methods

Analysis on livedet 2011

Analysis on livedet2013

Advantage

Cross – sensor Interoperability Less Error Rate as compared to existing methods

Thank You!