NonText Passwords

1
Non-Text Passwords

• CRyptography Applications Bistro
• Jessica Greer
• February 12, 2004

2
Outline

• Speech-Generated Cryptographic Keys
• Password Hardening Based on Keystroke Dynamics
• Other new ideas for non-text passwords based on
  behavioral biometric features

3
Key Generation

• Based on repeatable behavioral biometric
  characteristics
• timing
• force of keystrokes
• voice frequencies
• Aims to achieve two goals
• Breaking passwords will be no easier
• For some or most, breaking them will be harder

4
Speech-Generated Keys Monrose Reiter

• System initialization
• Generate key K
• Generate 2m shares of K using generalized secret
  sharing scheme, with m a system param
• Shares arranged within an m x 2 table such that K
  can be reconstructed from any set of m shares
  consisting of one share from each row

K
2
m
5
Twist on traditional secret sharing

• Traditional defense attacker will not possess
  enough shares to reconstruct the secret
• In this case, an attacker would have all shares
  if he had access to the physical device
• Requirement change that the attacker will not be
  able to find a sufficient set of valid shares in
  the table (make an exhaustive search
  computationally difficult)

6
Speech-Generated Keys Monrose Reiter

• Gathering behavioral measurements
• User utters passphrase
• System performs front-end signal processing and
  records measurements about voice features

My voice is my passport.
Verify me?
(photo from www.imdb.com)
7
Signal processing

• User utterance sampled at predefined sampling
  rate
• Minimum sampling rate on Compaq IPAQ 32 kHz
• Reduce computational and storage cost by down
  sampling to 8 kHz (sufficient to accurately
  capture signal) throw 3 of 4 samples away

8
Signal processing

• Signal then broken down and cleaned up
• Sample must be clean so as to be an accurate
  representation of users voice
• Arranged into frames 12-dimensional vectors of
  reals
• Background noise removed by calculating avg.
  noise in white space in the sample and
  subtracting it from entire length of sample
• Sample data converted to bit sequence called a
  feature descriptor used to regenerate key

9
Gathering behavioral statistics

• System measures m behavioral features of a users
  utterance
• Array of measurements concatenated into a bit
  string for each login attempt

10
Gathering behavioral statistics

• For each successful login attempt, the system
  updates the history of feature descriptors
  (consistent behavioral features)

11
Distinguishing features

• Security depends upon number of distinguishing
  features of voice
• A feature bai (a the account, i the feature) is a
  distinguishing feature if
• Ti gt avg(bai) - k stddev(bai) or
• Ti lt avg(bai) - k stddev(bai)

12
Going back to the 2 x m table

• Elements of table not consistently accessed are
  randomly perturbed
• Correct user should not encounter perturbed
  (invalid) elements in table
• The more often the user logs in, the stronger the
  system becomes

13
Empirical results

• For an implementation in which the table was also
  encrypted with a password makes a dictionary
  attack against the password up to 215 times more
  difficult

14
Password hardening based on keystroke dynamics

• Very similar concept system begins as secure as
  a traditional password system and begins
  perturbing values in secret-sharing table that
  are not repeated consistently

15
Potential problems

• Painful to change password, if security greater
  than traditional systems is essential cost
  associated with retraining the system
• In keystroke system, some degree of inference can
  be made about keystroke dynamics if password is
  known, and vice versa
• Not ideal for users who use different keyboards
• Security determined by degree of uniqueness of
  users voice or typing style

16
Is it accurate enough?

• Bergadano, Gunetti, and Picardi think not
• Inherent variability in most behavioral
  biometric identifiers is too great
• Propose using much longer samples and generating
  key based on duration of digraphs and trigraphs
  (sets of two and three consecutive letters)
• Not an appropriate substitute for traditional
  password systems
• Greater inherent variability with longer
  samples?

17
For more information

• www.biopassword.com
• Free demo
• www.mytec.com