Biometric Authentication Revisited: Understanding the Impact of Wolves in Sheep’s clothing in Sheep’s clothing Lucas Ballard, Fabian Monrose, Daniel Lopresti Presented by : Anuj Sawani 1
Biometric Authentication Revisited:
Understanding the Impact of Wolves
in Sheep’s clothingin Sheep’s clothingLucas Ballard, Fabian Monrose, Daniel Lopresti
Presented by : Anuj Sawani
1
Biometrics
• What is it?
– identifying, or verifying a person based on
• Physiological characteristics
• Behavioral characteristics
– Examples?– Examples?
• Biometric Authentication vs Identification
– “Am I who I claim to be?”
– “Who am I?”
• Better than passwords?
2
Handwriting as a biometric
• Offline
– 2-D bitmap
• Online
– Real-time data– Real-time data
• Signatures as a biometric?
Feature extraction Hash/Key
3
So, what’s with the menagerie?
• Sheep
– Easily accepted by the system
• Goats
– Exceptionally unsuccessful at being accepted– Exceptionally unsuccessful at being accepted
• Lambs
– Exceptionally vulnerable to imitations
• Wolves
– Exceptionally successful at imitations
4
The Threat Model
• Exploiting poorly protected template
databases
• Eavesdropping communication between • Eavesdropping communication between
sensor and the system
• Presenting artificially created samples to the
sensor
5
A neat idea – Concatenation attack
• Samples of user’s handwriting from other
contexts
• General samples of the style of writing
• Feature analysis …• Feature analysis …
• Generate the user’s handwriting synthetically!
6
Forgery styles
• Naïve
– Use other users’ writing as it was naturally rendered to forge the passphrase
• Naïve*• Naïve*
– Similar to Naïve, but uses similar writing styles
• Static
– Forgery using an image of the passphrase
• Dynamic
– Real-time rendering of the passphrase
8
Grooming the sheep into wolves
• 11,038 handwriting samples
• Incentives awarded to consistent writers,
“dedicated forgers”
• Three Rounds• Three Rounds
1. Collect the samples
2. Static and Dynamic forging
3. Selected “trained” forgers
9
Handwriting features
• How difficult is the feature to forge?
• Signals – t, x(t), y(t), p(t)
• For every feature f
– rf � missed by legitimate users– rf � missed by legitimate users
– af � missed by forgers
• Quality metric
– Q = (af - rf + 1)/2
• Q = 0 – never reliably reproduced by users
• Q = 1 – never reproduced by forgers
10
The winning features
• The probability that the ith stroke of c1
connects c2
• Median gap between the adjacent characters
• Median time between end of c and beginning • Median time between end of c1 and beginning
of c2
• Pen-up velocity
• A total of 36 good features out of 144
11
Algorithm to generate a known
passphrase• Select n-grams from different context such that
– g1 || g2 || … ||gk = passphrase
• Normalize t, x(t) and y(t) – match baselines
• Spatial adjustment of x(t)– Use median gap feature
• Fabricate p(t)• Fabricate p(t)– Use probability of connection feature
– Delayed strokes pushed into stack• Executed after each pen-up
• Add time delays– Use median time feature
– Use pen-up velocity and distance between strokes
12
The system at work…
• Used small sample set of 15 samples of user’s writing
– Each character from passphrase exists in set
– Does not include passphrase– Does not include passphrase
• Also, used 15 samples of similar writing style
• The algorithm caused an EER of 27.4%
– Forgers caused an EER of 20.6%
• n-gram length < 2
• Used 6.67 of the samples on average
13
Conclusion
• Handwriting as a reliable biometric?
– Refutable
• Adversary has been under-estimated till now
• Generative approach produces better • Generative approach produces better
forgeries than trained humans
14