Bypassing the Word Recognition Security Measure

1
Bypassing the Word Recognition Security Measure

• Isa M. Muqattash

2
Agenda

• Intro goal of project
• Project implications
• Difficulties expected
• Methods proposed time line

3
Introduction

• Many log-in sites use complex images containing
  text that is difficult to interpret.
• The main purpose is to cut down on security
  attacks by requiring live human interaction.
• Goal of project To design methods and implement
  a tool that is capable of recognizing the text of
  the image.

4
Project Implications

• Give insight into more strict requirements for
  image creation systems.
• Determine the strength of this security measure.
• We may give the bad guys an extra tool to use!

5
Difficulties Expected

• UNDERSTANDING HUMAN PERCEPTION
• Complex images that may even not be easily
  recognized by even humans. Good or bad ??
• Similarities between different characters ??
• Various fonts Colors
• Domain of text characters
• Many situations to consider...
• Pre-processing steps their order

c e !_at_ / .,
6
Preprocessing

• Cropping
• Remove colors
• GRAYSCALE... BLACK AND WHITE...
• Blurring Break down egde strength of additional
  add-ons
• ORDER OF OPERATIONS ??

7
Methods of Attack

• Character modelling
• Text recognition
• Area scanning Topological transformation

8
Character Modeling

• Build models for each character
• Proposed Methodology
• Edge detection
• Directional vectors
• Eliminate possibilities based on the directional
  vectors, until one possibility remains by
  comparing against a pre-build character model
  dataset.
• Straighten edges ???
• THIS METHOD TRIES TO FIT THE IMAGE INTO THE
  POSSIBILE CHARACTERS

9
Text Recognition

• Squeeze image along both axis to make text
  thinner
• Use traditional text recognition techniques
  and/or already-existing software
• Are we going to lose image information??
• The idea is that less important details are more
  likely to be thinner ???

10
Area Modeling

• Scan the image for a fixed anticipated character
• Topologically keep changing the sliding character
  to see if we may get a fit
• Use a threshold on matching area percentage to
  confirm acceptance or rejection of character
  scanned
• THIS METHOD FIXES THE ANTICIPATED CHARACTER AND
  SCANS FOR FOR IT IN THE IMAGE

11
Time Line

• One week for each method
• Character modelling
• Text recognition
• Area scanning Topological transformation
• One week to compare the different methods