“A Low-cost Attack on a Microsoft
CAPTCHA” By Jeff Yan and Ahmad Salah El Ahmad
The annual ACM Computer and Communications Security Conference (2008)
Presentation by Kathleen Stoeckle
Outline
• Overview on CAPTCHA
• Related Work
• The MSN CAPTCHA
• Microsoft CAPTCHA Segmentation
Attack
• Results & Analysis
• Strengths and Weaknesses
Overview of CAPTCHA
CAPTCHA
Completely Automated Public
Turing Test to tell Computers
and Humans Apart
Primitive CAPTCHAs
developed in 1997 by Andrei
Broder, Martin Abadi, Krishna
Bharat, and Mark Lillibridge.
Luis von Ahn and Michael
Blum coined the term
“CAPTCHA” and improved the
method in 2000.
http://www.searchenginepeople.com/blog/page/16http://www.searchengine
people.com/blog/page/16
Google CAPTCHA
http://www.codinghorror.com/blog/images/google-error-were-sorryrate-limiter-captcha.png
Yahoo CAPTCHA
http://www.carolisayazakuser.com/helppage/1aghostloginauthbox1a.gif
MSN CAPTCHA
http://www.msn.com
Text-Based CAPTCHAs
• Most widely used CAPTCHA
• Distort text images in order to make them
unrecognizable to pattern-recognition
programs.
• Popular because:
▫ Task is intuitive (character recognition)
▫ Few localization issues (roman characters
easily recognized)
▫ Strong security potential
Good CAPTCHAS
• Robust
• Human Friendly
Character Recognition
• Computers excel at recognizing characters, even
when the characters are distorted.
• If the positions of characters are known in
CAPTCHAs then breaking the scheme is a matter
of pattern recognition.
• When the positions are not known then the
computer programs have difficulty solving them.
Recognition rate for individual
characters under different distortions
Segmentation
• Segmentation – Identifying characters in the
right order.
• Challenging for both handwriting recognition
and computer vision.
• Traditionally, both computationally expensive
and difficult when taking into account all
characters in the challenge.
“State of the Art” CAPTCHAs
• Robustness of CAPTCHA must rely on
segmentation rather than recognition.
• If a text-based CAPTCHa is reduced to the
challenge of recognizing individual characters
then the scheme is effectively broken.
Purpose of Paper
Yan and El Ahmed’s paper examines the security
of the Microsoft CAPTCHA.
• This scheme was designed by an interdisciplinary
Microsoft team. The principle “segmentation resistance”
was established by this team.
• By attacking this CAPTCHA, the authors goal is to
determine how the MSN scheme and similar CAPTCHAs
can be improved.
• The paper shows how the Microsoft CAPTCHA was
broken with a desktop computer with a 1.86 GHz Intel
Core 2 CPU and 2 GB RAM using their algorithms.
Related Work
E-Z Gimpy and Gimpy
Broken by Mori and Malik:
• E-Z Gimpy (92% success)
• Gimpy (33% success)
• Object-recognition algorithms
http://www.cs.sfu.ca/~mori/research/gimpy
E-Z Gimpy and Gimpy
Broken by Moy et al:
• E-Z Gimpy (99% success)
• 4-letter Gimpy-r (78% success)
• Used distortion elimination
techniques
http://www.cs.sfu.ca/~mori/research/gimpy
Other Work…
• Chellapilla and Simard attacked visual
CAPTCHAs (4.89% to 66.2% success)
• Yan and El Ahmad defeated CAPTCHAs
generated on Captchaservice.org (Almost 100%
success).
▫ Accomplished by counting pixels of segmented
characters.
▫ Examined robustness from security angle.
▫ Simple pattern-recognition analysis.
• PWNtcha – a website that demonstrates
CAPTCHAs weakness and inefficiencies. Broke
visual CAPTCHas (49 to 100% success)
The MSN CAPTCHA
The MSN CAPTCHA Challenge
• Each challenge consists of 8 characters.
• Only upper case letters and digits are used.
• Text is dark blue and background is light
gray.
• Warping is used to distort characters.
• Random arcs of different thickness are
used in the anti-segmentation
measure.
Warping
• Local
▫ Small ripples, waves and elastic deformations
along the pixels of the character.
• Global
▫ Character-level, elastic deformations to foil
template matching algorithms.
Warping, cont’d
Local
Global
Random Arcs
• Thick Foreground Arcs
▫ Same color as characters
▫ As thick as the characters
▫ Non-intersecting
• Thin Foreground Arcs
▫ Same color as characters
▫ As thick as the thinnest parts of characters
▫ Intersecting
• Thin Background Arcs
▫ Thin
▫ Same color as background
▫ Cut through characters
Low-Cost Segmentation Attack
On Microsoft CAPTCHA
Low-Cost Segmentation Attack
•
•
•
•
Goal: Segment Microsoft CAPTCHA challenges.
Identify and remove random arcs
Identify all character locations in the right order.
Accomplishes this by:
▫ Dividing each challenge into 8 ordered segments.
Low-Cost Segmentation Attack
•
•
•
•
Goal: Segment Microsoft CAPTCHA challenges.
Identify and remove random arcs
Identify all character locations in the right order.
Accomplishes this by:
▫ Dividing each challenge into 8 ordered segments.
Low-Cost Segmentation Attack
•
•
•
•
Goal: Segment Microsoft CAPTCHA challenges.
Identify and remove random arcs
Identify all character locations in the right order.
Accomplishes this by:
▫ Dividing each challenge into 8 ordered segments.
Low-Cost Segmentation Attack
•
•
•
•
Goal: Segment Microsoft CAPTCHA challenges.
Identify and remove random arcs
Identify all character locations in the right order.
Accomplishes this by:
▫ Dividing each challenge into 8 ordered segments.
Attack in 7 Steps
• Binarization
• Fixing Broken Characters
• Vertical Segmentation
• Color Filling Segmentation
• Thick Arc Removal
• Locating Connected Characters
• Segment Connected Characters
Step 1: Binarization
1. Convert a color challenge to a two-color image
using threshhold method.
2. High intensity  White
3. Low intensity  Black
Step 2: Fixing Broken Characters
1. Keep character as a single entity.
2. Prevent small portions of characters from being
removed by an arc.
Step 2
• Find background color pixels that have left and
right neighbors with foreground color
• Find background color pixels that have top and
bottom neighbors with foreground color.
• Convert pixels identified above to foreground
color.
Step 2
Step 2
Step 3: Vertical Segmentation
Segmentation method – Divide challenge
vertically into chunks.
Divide and Conquer
Step 4: Color Filling Segmentation (CFS)
• CFS applied to each chunk (Step 3)
• Find every connected component or “object” in
each chunk.
Step 4
• CFS Algorithm:
1. Detect foreground pixel and trace it to all
connecting pixels. This creates an object.
2. Locate foreground pixel outside of the
object and traces connecting pixels to
identify the next object.
3. Process essentially amounts to color filling
each object. The number of colors used =
the number of objects.
Step 4
8 connectivity - Each pixel has 8 neighbors
Step 4
A color fill is applied to each chunk,
regardless of number of objects in the
chunk.
Step 5: Thick Arc Removal
Thick Arc Characteristics:
1. Pixel Count – Generally small
2. Location – Close to or intersecting with image
border.
3. Shape – Thick arcs do not contain circles. No
characters such as A, B, D, P, Q, 4, 6, 8 and 9.
4. Interplay between shape and location –
Correlation between thick arcs and geometric
shape:
•
•
Tall but Narrow near start of CAPTCHA
Wide and Short near middle
Step 5
Thick Arc Removal Algorithm
1. Circle Detection
2. Scan objects without circles for
distinct features.
3. Relative Position Checking
4. Detection of Remaining Arcs
Step 5-1: Circle Detection
1. Draws bounding box around an object.
2. Fill box with a color that is different
from foreground and background.
3. Scan for pixels with background color.
If found, a circle has been detected.
Step 5-2: Scan non-circle objects for
distinctive features
Pixel Checking
 Characters generally have pixel count of
over 50.
 Any characters 50 pixels or less is
removed as an arc.
Step 5-3: Relative Position Checking
This step is applied to all chunks with more than one object.
Premise: The positions of objects determines the
difference between arcs and characters. Characters are
always closer to the baseline. Characters are horizontally
juxtaposed, but never vertically.
Step 5-4: Detection of Remaining Arcs
▫ Count the number of remaining arcs in
the image.
▫ Remaining arcs are generally the first and
last object in the current image.
▫ Check first and last objects using these
rules:
1) If one object contains circle, the other is
removed.
2) If neither object contains a circle, the one
with the fewer number of pixels is removed.
Step 6: Locating Connected Characters
n = number of objects in an image
If n< 8, at least one object has two or more connected
characters.
MSN Challenge
1. 8 characters in an image
2. Connected characters are connected
horizontally not vertically and thus are
wider.
3. A segmented chunk contains more than one
character if the chunk is wider than 35 pixels.
The number of chunks, width of chunks, and number of objects
in a chunk are used to guess which chunks contain
connected characters.
Step 6: Locating Connected Characters
Step 7: Segment Connected Characters
•
•
Find the width of an object by determining its left
and rightmost pixels.
Vertically divide object into c parts of same width
where c = number of characters.
Results & Analysis
Results of 7 Step Attack
Success Rate: 91% (91 out of 100 challenges)
92% of 500 random challenges
Attack Speed:
Implemented in java
1.86 Ghz Intel Core CPU and 2 GB Ram
Implications: A “state of the art” machine can
achieve at least a 95% success rate for recognizing
individual characters in MSN scheme. This is a
conservative estimate.
Overall success rate for breaking the MSN
CAPTCHA: 61% (≈ .92*.95^8).
Problems with the Attack
1. Failure of Arc Removal
2. Failure of Approximation
3. Failure of Segmentation of Connected Characters
Arc Removal and “Approximation” = 72.8%/82.5%
Analysis of Failed Cases
Defense Against this Attack
• Let characters touch/overlap
• Juxtapose characters in any direction in
order to make it harder to tell characters
and arcs apart.
• Randomly vary the width of the characters
Strengths and Weaknesses of MSN scheme
Strengths:
Good usability – characters generally
recognizable
Weaknesses:
Security – vulnerable to simple segmentation
attack.
Usability – many characters are still
unrecognizable.
Considerations
Size matters
Both character size and string size. The longer the
string, the more security it provides.
String Length
•
If random characters are used, the longer the
string, the lower the usability.
•
If dictionary words are used, the longer the
string, the greater the usability.
Fixed Length
Aids segmentation attack, but can improve usability.
Conclusion?
Strengths and Weaknesses
Strengths
• Paper explained how the attack worked very clearly.
• Good analysis of both successes and failures of the
attack.
• Included area of future study which is useful.
Weaknesses
• Organization
• Despite the clear explanation, the paper included
no insight on how the authors created the algorithm.
• Much of the attack was based on the premise that
the CAPTCHA has 8 letters. With the evolution of
CAPTCHAs and attacks, this might generate an idea
of false security if CAPTCHAs have a random
number of letters and/or have less applicable for
analyzing other CAPTCHA schemes.
References
• A Low-cost Attack on a Microsoft CAPTCHA, Jeff Yan, Ahmad Salah El
Ahmad, CCS 2008. http://homepages.cs.ncl.ac.uk/jeff.yan/msn_draft.pdf
• http://en.wikipedia.org/wiki/Captcha
• http://www.searchenginepeople.com/blog/page/16http://www.searchengi
nepeople.com/blog/page/16
• http://www.codinghorror.com/blog/images/google-error-were-sorry-ratelimiter-captcha.png
• http://www.virtualblight.com/articles/wpcontent/uploads/2008/03/captcha_1_4.jpg