Brain W ave B ased A uthentication

Brain Wave Based Authentication

Kennet Fladby2008

Outline

1. Introduction2. Research questions3. Experimental work4. Results5. Conclusion6. Further work

1-1. Brain waves

• The brain contains about 100 billion neurons.• Neurons generates and leads electrical signals.• The sum of these electrical signals generates

an electric field.• Fluctuations in the electric field can be

measured.• Electroencephalographic (EEG)

1-2. 10-20 System

1-3. EEG signal: 20 seconds, 128Hz

2. Research questions

• Is it possible to authenticate by means of brain waves with only one EEG sensor?

• What feature should be extracted from the signals?• Do we have to authenticate based on a person’s

thoughts or can we use the brain waves as a biometric directly?

• Will a distance metric approach work?• What is the best FMR and FNMR we can achieve?

3-1. TasksTask Description

Relax Relax in a normal fashion

Color Visualize the red color

Rotate Mentally rotate a house

Password Think about the password ’BrainWaveS’

Music Think about a melody/song

Words Generate words with capital letter ’M’

Count Count upwards starting from 1

Read Read a random provided text

3-2. Setup

• 10 participants– 3 sessions, 3 recordings of each task per session– Each recording lasts 20 seconds (2560 samples)– Eyes closed

• Number of recordings– 72 per participant ( 24 minutes )– 720 total (4 hours )

3-3. Physical movement anomalies

3-4. Initialization problem

3-5. Frequency domain• The brain operates at low frequencies usually

divided into six frequency bands:Frequency band Range

Delta 1 – 4Hz

Theta 4 – 8Hz

Alpha 8 – 12Hz

Beta-Low 12 – 20Hz

Beta-High 20 – 30Hz

Gamma 30 – 50Hz

3-6. Fast fourier transform

3-7. Feature extraction

• Time domain features– Mean sample value– Zero crossing rate– Values above zero

• Frequency domain features– Peak frequency– Peak frequency magnitude– Signal power in each frequency band

• Pdelta, Ptheta, Palpha, PbetaLow, PbetaHigh, Pgamma

– Mean band power– Mean phase angle

3-8. Statistics

• Chi-square goodness-of-fit test– Samples and features do not follow normal

distribution.

• Correlation– High correlation between PbetaLow and

PbetaHigh (8 out 10 participants).

3-9. Distance metricd = d(signal1,signal2) :

X = signal1Y = signal2

d1 = |X.PbetaLow / X.PbetaHigh - Y.PbetaLow / Y.PbetaHigh|d2 = |X.PbetaLow / Y.PbetaLow - Y.PbetaHigh /X.PbetaHigh|d3 = |X.Palpha / X.PbetaLow - Y.Palpha / Y.PbetaLow| d4 = |X.Palpha/ Y.Palpha - Y.PbetaLow / X.PbetaLow|

d = d1 + d2 + d3 + d4

4-1. Distance computation 1

• Computation: All vs All• Genuine attempts:– d(signal1,signal2) from the same participant

• Fraudulent attempts– d(signal1,signal2) from different participants

• Requirement:– d(signal1,signal2) must be from the same task

4-2. DET-Curve 1

EER = 30.28%

4-3. Distance computation 2

• Computation: All vs All• Genuine attempts:– d(signal1, signal2) from the same participant

• Fraudulent attempts– d(signal1, signal2) from different participants

• Requirement:– d(signal1, signal2) must be from the same task

AND the same session.

4-4. DET-Curve 2

EER = 23.26%

4-5. Task selectionTask with the best average distance

Participant Session 1 Session 2 Session 3

1 Color Count Words2 Count Count Password3 Color Count Rotate4 Words Color Rotate5 Rotate Password Password6 Count Count Words7 Rotate Words Color8 Rotate Password Words9 Words Words Music10 Relax Color Rotate

4-6. Distance computation 3

• Computation: Task selection• Genuine attempts– d(signal1,signal2) from the same participant

• Fraudulent attempts– d(signal1,signal2) from different participants

• Requirement– d(signal1,signal2) must be from the selected

session 1 task.

4-7. DET-Curve 3

EER = 21.46%

4-8. Distance computation 4

• Computation: Task selection• Genuine attempts– d(signal1,signal2) from the same participant

• Fraudulent attempts– d(signal1,signal2) from different participants

• Requirement– d(signal1,signal2) must be from the selected

session 1 task AND the same session.

4-9. DET-Curve 4

EER = 17.08%

4-10. DET-Curve 1-4

EER = 30.28%EER = 23.26%EER = 21.46%EER = 17.08%

5. Conclusion

• Similiarities are session based– Two consequtive signals are very similar

• Equipment dependant– Signal gets better over time– Captures too much physical movement

• One sensor is not enough– Limited information– Low sample rate

6. Further work

• Better distance metric– Identify more feature relations– Try different feature combinations

• Better selection of tasks– Tasks designed for the Fp1 location

• New equipment– Better filtering– Increased sample frequency– More sensors– Different sensor locations

Thank you for listening!

Questions?