Faces in the Distorting Mirror: Revisiting Photo-based Social Authentication

Faces in the Distorting Mirror:
Revisiting Photo-based Social Authentication
Iasonas Polakis
Dept. of Computer Science
Columbia University
New York, NY, USA
polakis@cs.columbia.edu
Panagiotis Ilia
Institute of Computer Science
FORTH
Heraklion, Greece
pilia@ics.forth.gr
Federico Maggi
DEIB
Politecnico di Milano
Milan, Italy
federico.maggi@polimi.it
Marco Lancini
CEFRIEL
Milan, Italy
marco.lancini@cefriel.com
Georgios Kontaxis
Columbia University
New York, NY, USA
kontaxis@cs.columbia.edu
Stefano Zanero
DEIB
Politecnico di Milano
Milan, Italy
zanero@elet.polimi.it
Sotiris Ioannidis
Institute of Computer Science
FORTH
Heraklion, Greece
sotiris@ics.forth.gr
Angelos D. Keromytis
Columbia University
New York, NY, USA
angelos@cs.columbia.edu
ABSTRACT
In an e↵ort to hinder attackers from compromising user
accounts, Facebook launched a form of two-factor authen-
tication called social authentication (SA), where users are
required to identify photos of their friends to complete a
log-in attempt. Recent research, however, demonstrated that
attackers can bypass the mechanism by employing face recog-
nition software. Here we demonstrate an alternative attack
that employs image comparison techniques to identify the
SA photos within an o✏ine collection of the users’ photos.
In this paper, we revisit the concept of SA and design
a system with a novel photo selection and transformation
process, which generates challenges that are robust against
these attacks. The intuition behind our photo selection
is to use photos that fail software-based face recognition,
while remaining recognizable to humans who are familiar
with the depicted people. The photo transformation process
creates challenges in the form of photo collages, where faces
are transformed so as to render image matching techniques
ine↵ective. We experimentally confirm the robustness of our
approach against three template matching algorithms that
solve 0.4% of the challenges, while requiring four orders of
magnitude more processing e↵ort.. Furthermore, when the
transformations are applied, face detection software fails to
detect even a single face. Our user studies confirm that users
are able to identify their friends in over 99% of the photos
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than the
author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from permissions@acm.org.
CCS’14, November 3–7, 2014, Scottsdale, Arizona, USA.
Copyright is held by the owner/author(s). Publication rights licensed to ACM.
ACM 978-1-4503-2957-6/14/11 ...$15.00.
http://dx.doi.org/10.1145/2660267.2660317.
with faces unrecognizable by software, and can solve over
94% of the challenges with transformed photos.
Categories and Subject Descriptors
K.6.5 [Management of Computing and Information
Systems]: Security and Protection—Authentication
General Terms
Security, Human Factors
Keywords
Social Authentication; Image Analysis; Face Recognition;
CAPTCHAs
1. INTRODUCTION
The abundance of personal information uploaded to online
social networks (OSN), coupled with the inherent trust that
users place into communications received from their contacts,
has rendered compromised profiles a lucrative resource for
criminals [23]. Moreover, the widespread adoption of single-
sign on services o↵ered by popular OSNs, makes user profiles
even more valuable. Consequently, researchers designed var-
ious systems for detecting compromised accounts [10, 13].
However, preventing unauthorized access in a user-friendly
manner remains an open problem. To safeguard profiles
against attackers that have stolen user credentials Face-
book deployed a countermeasure called social authentication
(SA) [21]. This is basically a variant of the traditional two-
factor authentication scheme (e.g., [4, 5]), which requires
users to identify their contacts in a series of photos.
Social authentication is a promising approach, as it o↵ers
a user-friendly mechanism to strengthen the login process.
Researchers, however, have analyzed [16] its weaknesses, and
demonstrated [20] that the existing system is vulnerable to
attacks that employ face recognition software. We further

demonstrate that SA is vulnerable to an attack that previous
work has overlooked; the adversary first builds a collection of
the photos uploaded by the victim and his online friends. The
adversary can then solve the challenges by identifying the
photos within the collection via image comparison, and using
the tag information to select the correct answer. Compared to
the previous attack, this attack has an important advantage:
the identification of photos within a collection based on
image comparison techniques is far more accurate than face
recognition, and e↵ective even when no faces are present.
In this paper, we revisit the concept of SA and build a
system that retains the usability of the existing mechanism,
while being robust to attacks employing image analysis soft-
ware. We conduct a user study that provides us with valuable
information regarding a critical aspect of our approach to
SA; the ability of users to identify their friends in photos
taken under realistic, non-ideal conditions. The participants’
photos are processed by state of the art face recognition
software and categorized as “simple” , “medium” or “di cult”,
based on the quality of the faces found (if any). While SA
picks simple photos, we focus on the medium and di cult
categories. Users solve over 99% of the medium and 82%
of the di cult challenges, indicating their ability to identify
their friends even when their faces are not clearly visible,
based on secondary features (e.g., posture, hair), associative
information (e.g., pets, objects) or memory retrieval (users
remember having seen the photos). On the other hand, face
recognition software fails to identify the users in such photos.
Analysis of the results of the user study and the charac-
teristics of the two attacks provides significant insight that
allows us to design a secure, yet usable, system that ren-
ders the attacks ine↵ective. First, we redesign the photo
selection procedure, which processes photos with face recog-
nition software, and selects those that contain faces but are
not recognizable by face recognition software (i.e., medium
photos). Next, we apply a novel transformation process for
creating the SA challenges, which hinders image comparison
techniques from mapping them to the original photos.
Our prototype implementation creates SA challenges by
superimposing the selected medium faces (tags) over the faces
of a random “background” photograph. The overlayed faces
are made transparent so as to blend in with the underlying
faces. Then, a perspective transformation is performed on the
photo, which prohibits even highly resilient pattern matching
approaches, like template matching, from mapping them to
the original photos.
Subsequently, we conduct an extensive set of experiments
using real data, to measure how various levels and combina-
tions of the transformations impact the attacks. The results
demonstrate the robustness of our challenges, as our system
completely hinders the face recognition attack that fails to
detect any faces. The challenges are even robust against
the image comparison attack that employs three template
matching algorithms; all three pass less than 2% of the chal-
lenges with two tags, and 0.4% of those with three tags, while
requiring four orders of magnitude more processing e↵ort
than against the non-processed photos.
To verify that people depicted in the photos remain iden-
tifiable after the transformations, we conduct a preliminary
user study where users are asked to identify famous people
in a series of challenges. Results verify the usability of our
system, with users solving 94.38% of the challenges.
Finally, we discuss the applicability of our approach as
a security service o↵ered by an OSN to other websites. In
our vision, this mechanism can be adopted by web services
as a user-specific CAPTCHA service, or even by high value
services (e.g., banking websites) as a security measure addi-
tional to two-factor authentication. We discuss its robustness
against attacks that break traditional CAPTCHAs, like out-
sourcing attacks, and argue that it is a user-friendly and
secure alternative to existing schemes.
In summary, the key contributions of this work are:
• We demonstrate a novel attack technique against social
authentication, that is more e↵ective and e cient than
the one previously presented.
• We conduct the first user study that explores the ability
of humans to identify their acquaintances in photos
taken under realistic, non-ideal conditions. Our results
demonstrate that humans solve this task e↵ectively,
even when no faces are in the photos.
• Based on the insights derived from our experiments, we
design a secure, yet usable SA mechanism, that relies
on a novel tag selection and transformation process. We
experimentally evaluate our proof-of-concept implemen-
tation which completely hinders the face recognition,
and reduces the success of the image comparison attack
to 0.4%, while requiring four orders of magnitude more
processing e↵ort from the attacker.
• To verify the usability of our system, we conduct a
preliminary user study where users solve 94.38% of the
challenges with transformed photos.
2. ATTACKING SOCIAL AUTHENTICATION
Photo-based authentication in OSNs was first presented
in 2008 by Yardi et al. [27]. In 2010 Facebook deployed its
SA application in an e↵ort to prevent adversaries from using
stolen credentials. In a nutshell, when a login attempt is
considered suspicious, the system presents the user with a
series of 7 pages, each containing 3 photos of a friend and 6
potential answers. The user is required to correctly identify
the friends depicted in at least 5 of the pages.
Face Recognition Attack. In previous work we demon-
strated that practically anybody can solve SA challenges, by
collecting publicly available data and employing o↵-the-shelf
face recognition software [20]. The photos and their tags
are used to train face recognition classifiers, which can iden-
tify the friends depicted in SA challenges. Our estimations
showed that 84% of Facebook users are susceptible to this
attack.
Image Comparison Attack. Attacking SA needn’t rely
on face recognition, as more e↵ective photo matching tech-
niques can be used instead. The attacker first creates a
collection with all the victim’s friends’ photos he can access,
along with the tag information. When SA is triggered, the
attacker identifies the presented photos within the collection,
and uses the tag information to answer the challenge.
The advantage of this attack is its e↵ectiveness even when
the challenges contain faces that cannot be identified via
face recognition. Regardless of the content the adversary can
pass the challenge if some of the photos are in the collection
(at least 1 of the 3 photos, in 5 of the 7 pages). The attack
success is proportional to the coverage of photos. This can

Collection size 5K 10K 20K 30K 40K
Identified photos 98.8% 98.4% 98.4% 98.4% 98.4%
Table 1: Identified photos in image comparison at-
tack, for di↵erent collection sizes.
(a) Simple (b) Medium (c) Di cult
Figure 1: Sample photo from each category.
be increased by employing fake accounts to befriend the
victim’s friends. However, publicly available photos are a
good resource for the attack. According to [18], the authors
found that 63% of the photos have privacy settings di↵erent
from what the users intended. Alarmingly, the privacy setting
for 51% of those photos was set to public, allowing anyone
to access them. Thus, overall, about 1 out of every 3 photos
will be publicly viewable by accident. Nonetheless, previous
work (e.g., [6, 7, 14, 25]) has extensively demonstrated the
e↵ectiveness of employing fake accounts to befriend users of
OSNs, and have reported success rates of up to 90%.
Various image comparison techniques can be used for this
attack. Here we demonstrate its e↵ectiveness even when
employing a simplistic pixel comparison. We assume the
attacker has access to all of the user’s photos (we retain
the same assumption when defending against this attack in
Section 4.3). Experiments were conducted on a 4-core IntelR
Core
TM
i7-4770 CPU @ 3.40GHz equipped with an SSD.
We build collections of varying sizes, and create 100 SA
challenges from each collection. The collections are up to
40,000 photos, which is higher than the average from our
user study (Section 3.2). We employ a simple and fast
image comparison approach to identify the presented photos
within our collection: we crop the top left corners of the
presented photos, and match them to those of the photos
in the collection. In certain cases this might result in false
positives (e.g., the top left corner of the photo has only black
pixels), however it did not a↵ect our success rate. Table 1
presents the results of our experiments, indicating the average
ratio of photos identified correctly in each SA challenge (21
photos). We passed all the challenges with at least 18 of the
photos identified, and had identified at least 98.4% of the
photos in all scenarios.
Our approach is very e cient as we are able to identify
the 21 photos within 40K photos in 1.24 seconds (⇠ 0.06
per photo). One could possibly improve performance by
intersecting the suggested names and the tags after each
photo identified within a page.This could decrease times, as
the tags from one photo might be enough to infer the answer.
3. MEASURING USER ABILITIES
To design a secure SA system that exploits noisy and
unidentifiable faces, we need to verify our intuition that
humans are capable of recognizing their friends in photos
taken under natural conditions. Although previous work [12,
22] has explored the ability of people to discern human faces
or their features, we are the first to focus on the ability
of recognizing friends (as opposed to unknown faces), even
under conditions where the faces may not be clear or even
present at all.
Measurement Application. We created a Facebook
app that replicates the SA mechanism, which require users
to identify their friends in SA challenges, and complete a
questionnaire for each photo. We opted for a Facebook app
for two reasons: first, they inspire trust in users as they are
deployed within a sandbox and are governed by a series of
permissions that clearly state the user data accessed. Second,
a Facebook app enables direct access to some user profile
data (e.g., their social circle). This enables us to respect
user privacy and minimize collection of potentially sensitive
information, since we use data stored on Facebook rather
than having users upload it to our own infrastructure.
IRB Approval. We issued an IRB protocol request to
the review board of our institution, that clearly described the
parameters of our study, and the data we would be gathering.
After it was approved we invited users to participate.
Recruiting Users. We explored and experimented with
the possibility of reaching human subjects through the Ama-
zon Mechanical Turk (AMT) service [1]. However, asking
Turks to install an app, or directing them to a URL outside
Amazon to fill out a survey, explicitly violates the AMT
terms of services. Our tasks were rejected by Amazon be-
cause of this purely technical incompatibility. The nature of
our system, where challenges are crafted specifically for each
user, prohibited us from taking advantage of such crowd-
sourcing platforms. Therefore we resorted to recruiting users
directly by sharing our app with the online contacts of our
personal Facebook accounts, and posting flyers around the
university campus. We also o↵ered prizes as an incentive for
user participation. This allowed us to collect and analyze
a significant amount of user data, regarding over 4 million
photos, and over 1,000 solved SA challenges.
3.1 Measurement Workflow
Once a user installs our app, it collects the social graph
and related metadata (i.e., friends, URLs of photos, tags). It
processes the collected photos with state of the art face recog-
nition software, and categorizes them as simple , medium or
di cult , based on the quality of the faces found. Photos of
each category are selected to create challenges of increasing
di culty and measure the user’s ability to solve them.
Step 1: Face Extraction. We use the face.com online
service, which has since been acquired by Facebook [3]. Its
e↵ectiveness when using photos taken under realistic con-
ditions has been demonstrated [20], and it performs better
than other state of the art solutions [24]. We focus on two
specific metrics assigned to the detected faces.
Confidence: when detecting faces, face.com’s classifier
returns its level of confidence that the tagged area contains
a face. Tags assigned a low confidence level have a high
probability of not containing a face.
Recognizable: not all faces are suitable candidates for train-
ing (or being recognized by) a classifier: face.com returns a
boolean value to indicate this; “true” when faces can be recog-
nized or are suitable to be used as part of a training set, and
“false” otherwise. Even if a face is present, due to various rea-

Type Total Passed Success Per User
Simple 362 358 98.89% 3.98
Medium 347 344 99.14% 3.81
Di cult 335 275 82.09% 3.68
Total 1044 977 93.58% 11.47
Table 2: Number of challenges taken from each cat-
egory, and percentage of successfully passed ones.
sons (e.g., angle, obstacles) proper face-classification features
(e.g., haar, eigenfaces, fisherfaces) cannot be extracted.
Step 2: Photo Categorization. Based on these metrics,
our app assigns photos to the following categories:
Simple - Figure 1(a): photos containing tags that most
likely frame a human face. This is our baseline category,
as it replicates the existing SA mechanism, and provides
a reference for comparison. According to [20], 80% of the
photos presented in SA challenges by Facebook had a face in
the tagged area that was detectable by software. Therefore,
we select photos with high confidence ( 80%) which have
been classified as recognizable (recognizable=T).
Medium - Figure 1(b): photos with a high probability of
containing a face (confidence 80%), which are classified as
bad candidates for training/recognition (recognizable=F).
Di cult - Figure 1(c): photos classified with a confidence
below 40%. This category is to measure how e↵ective people
are at recognizing their friends even if their face is not visible.
Step 3: Photo Description. After a user selects the
name of each depicted friend, our app informs them if they
were right or wrong, and requires them to answer 4 questions
per photo describing: the content, the position and visibility
of the user’s face and other faces within the tagged area, and
reasons why the photo was useful or not (see Appendix A).
3.2 User Study Results
Our goal is to measure the users’ ability to recognize their
friends, and demonstrate that humans can solve this task in
conditions where the automated attacks would fail, as we
demonstrate in Section 4.3.
Collected dataset and demographics. 141 users in-
stalled our app which led us to a total of 4, 457, 829 photos
and 5, 087, 034 tags. However, 90 of the users actually com-
pleted challenges, out of which 79 were listed as male and 11
as female, from 6 di↵erent countries. Of the 82 that reported
their age, 63 were between 20 and 30 years old and 15 were
between 30 and 40. On average, users had 347 friends each.
Recognizing Friends. Table 2 presents the number of
challenges (each has 7 pages with 3 photos of the same user
and 6 suggested names) per category, and the percentage
that users were able to solve (recognize at least 5 out of
7). Results are surprisingly high and consistent for medium
challenges, as users solve over 99% of the challenges.Thus,
even for photos with faces that cannot be identified by state of
the art face recognition software, users’ success rates are not
impacted. Users also score surprisingly well for the di cult
challenges, with an 82% success rate.
Influence of the Social Circle Size. Figure 2 shows
the number of friends that a user has and the success rate for
solving SA challenges. Each point refers to the overall success
rate of a user for all 3 categories, and the label indicates
Number of friends
100 200 300 400 500 600 800 1000 1500
SuccessfulTests(%)
70
80
90
100
823 11 3 12123
8
2 3 111
10
1
9
3
29
3234
30
11
6
77
3
21
30
12
42
22
11812212
22
9
418 3
57
1
28
1185
9
3
9
12
69
21
9
7
311 11
21
151465
21
1361 9
6
18
9
6
Figure 2: Correlation between number of friends
and challenges solved. Each point’s label indicates
how many challenges the user has taken.
the total number of challenges that the user completed. As
the number of friends increase, we expect users to score
significantly lower. However, the results do not demonstrate
such an e↵ect, and no strong correlation is evident. The
suggestion of names is important, as users can decide based
on content that can be associated to certain friends. This
result is very encouraging, as it disproves concerns regarding
the applicability of SA for users with a large number of
friends. Here we only visualize users that have completed at
least 3 challenges for reasons of visual clarity, without the
overall distribution being a↵ected in any way.
Photo Content. Figure 3(a) shows the distribution of
answers regarding the content of the photos, which also o↵ers
an evaluation of the quality of the photo-detection process.
As expected, for the simple and medium categories, the
vast majority of photos (over 80%) are labeled as portraits,
meaning that the focus of the photos are human faces. In
contrast, in the di cult one they account for 37%. These
numbers verify how accurate the face detection process of
face.com is, as the confidence levels we have set in our photo
categorization process are verified by users. The di cult
category is more evenly distributed.
Face Position. Figure 3(b) plots the distribution of
answers about the placement of the friend’s face with respect
to the tagged area. 49% of the medium photos contain a
clearly visible face inside the tagged area (InClear) and an
unclear face in 27.8% (InUnclear), cumulatively approaching
the 80% confidence criteria. The di cult photos do not
contain the friend’s face (Absent) in about half the photos.
Presence of Other Faces. Figure 3(c) shows the distri-
bution of other faces being visible in the photo, and their
placement in regards to the tagged area. The simple and
medium categories contain some other face in 83% and 77.5%
of the photos with faces being outside the tag in 41% and 45%
of the cases respectively. For the di cult category, 43.5% of
the photos contain no human faces (Nobody).
Usefulness of the Photo. Figure 3(d) plots the distri-
bution of photos regarding their usefulness. The selected
friend was present in about 70% of the simple and medium
photos, which is less than the photos containing the friend’s
face according to Figure 3(b). This is due to users selecting
other options such as “remembering the photo” or “relevant
to this friend”, even though the friend’s face is in the photo.
An interesting aspect of the di cult category, where photos
have a low probability of containing the face, is users relying
on other information to correctly select the friend. This cat-
egory has a higher percentage of answers that rely on other

Portrait
Landscape
Objects
Animals
Text
Art
Photos(%)
0
10
20
30
40
50
60
70
80
90
100
simple
medium
difficult
(a) Content of photo.
InClear
OutClear
InUnclear
OutUnclear
Absent
Photos(%)
0
10
20
30
40
50
60
70
80
90
100
simple
medium
difficult
(b) Position and clarity of
friend’s face.
In Out
InOut
Only
Nobody
Photos(%)
0
10
20
30
40
50
60
70
80
90
100
simple
medium
difficult
(c) Position and clarity of other
faces.
InPhoto
Remember
Relevant
NoOneElse
NotUseful
Photos(%)
0
10
20
30
40
50
60
70
80
90
100
simple
medium
difficult
(d) Usefulness of photo.
Figure 3: Distribution of answers given by the users in our study.
Type Portrait Landscape Objects
Simple 97.4% (1133) 94.9% (59) 0% (1)
Medium 97.6% (1225) 90% (30) 0% (0)
Di cult 92.1% (267) 76.9% (26) 64.5% (31)
Table 3: Success rates (and total number) for pages
where all 3 photos were labelled as the same type.
InPhoto
NoOneElse
NotUseful
Relevant
Remember
Photos(%)
0
10
20
30
40
50
60
70
80
90
100
Text
Animals
Art
Objects
Landsc.
Portrait
Figure 4: Correlation between content of the photos
and usefulness as reported by users.
types of visual clues and context for excluding (NoOneElse)
or inferring (Relevant) suggested names.
Absolute Success Rate per Category. Table 3 shows
the statistics for pages (each SA challenge contains 7 pages
with 3 photos each) in which all 3 photos are assigned to the
same category. We present the percentage of pages in which
the depicted friend was correctly identified, and the total
number of pages in parentheses. For di cult portraits, users
were able to identify their friends in 92.1% of the pages. Thus,
people can identify their friends in photos where software
cannot even detect a face. In the medium category, users
were successful in more than 97% of the pages, validating
our initial intuition and the applicability of our approach.
Even though the number of the remaining di cult challenges
is too small to draw concrete conclusions, it is surprising
that success rates are over 64% in all cases, and users even
identified their friends in 77.7% of the pages that contained
photos of animals. Thus, associative correlations assist users
when presented with a photo of objects or pets.
Absence of Friend’s Face. To verify the ability of
users to infer the answer even when the user is not present,
Figure 4 breaks these numbers down. It becomes evident,
as the cumulative ratios for the Landscape, Objects, Text
and Art photos account for 44% and 55.5% of Relevant and
Remember respectively. Thus, almost half of the photos for
which users relied on inference or memory, are not of faces.
Remember_F
Remember_O
Relevant_F
Relevant_O
NoOther_F
NoOther_O
Photos(%)
0
10
20
30
40
50
60
70
80
90
100
OnlyFriend
Others
Nobody
Clear
Unclear
Absent
Figure 5: Breakdown of the photos that were useful
for inferring or excluding answers, in regards to the
friend’s face ( F) and other people’s faces ( O).
We then focus on photos where the depicted friend’s face
was absent. When the users remember the photo, they are
almost always able to correctly identify the friend (only one
case with a wrong answer). Surprisingly, users achieve high
success rates when selecting the depicted friend based on the
relevance of the content, being 97.4% for the di cult photos.
When users try to exclude suggested friends based on the
content until they can select the correct answer (NoOneElse),
they still achieve very high accuracy (albeit lower than the
Relevant ones) with a combined 89.6% across the 3 di culty
categories. As expected, when the photo is not considered
useful the success rates are lower. However, in the simple
and medium categories, these photos belong to pages that
were correctly answered in more that 82% of the cases, due
to the high probability of the other photos containing a face.
Total absence of faces. We wanted to explore whether
the existence of other people in the photo might influence the
results, i.e., even if the requested friend is not in the photo
but other common friends are, the user might still be able
to infer the correct answer. When we focus on photos where
the friend was absent, and no other faces were contained
either, the results remain almost identical. However, the
simple and medium photos that were flagged as NotUseful
present a measurable decrease (6.5%, 16.3%). As can be seen
in Figure 3(c), these categories have a much higher ratio
of photos that contain other faces compared to the di cult
category. Thus, while photos might not contain any faces,
the content can assist the user in inferring the correct answer
or excluding suggestions until left with a plausible one.
Exclusion and Inference. To explore the e↵ect of other
people’s faces on users excluding suggestions or inferring the
answer, we plot Figure 5, which provides interesting insight
on the strategy users follow for identifying the depicted friend
based on the presence of the friend’s (columns with F) and
other users’ (columns with O) faces. Users tend to remember

the photo when the depicted friend’s face was either Unclear
or Absent (Remember F), as is the case for users inferring the
correct answer (Relevant F). Users also tend to remember
photos where other people (common friends) or no people
are present at all (landscapes, pets and important objects).
Furthermore, in Relevant O we can see that users infer the
correct friend almost equally from photos where Nobody
(47.9%) is present (due to relevant objects and pets) or
where Other (44.5%) people are present (people that the user
knows are friends with the requested friend).
When excluding suggestions (NoOther F), the absence of
the friend’s face (Absent) or its poor quality (Unclear) have
a similar e↵ect. However, the presence of other people’s
faces has a major impact, as it accounts for 61.4% of the
photos in NoOther O. This is a strong indication that when
users are presented with photos of unknown individuals they
tend to follow an approach of excluding suggested friends
until left with a plausible answer. If the users know the
depicted people are acquaintances of the requested friend,
they select Relevant. In the case of unknown individuals,
they exclude suggestions of close friends and select less known
contacts that have a higher probability of being friends with
people unknown to the user. Combined with the information
that can be extracted from the other photos contained in a
challenge page, this approach can be very e↵ective, as users
correctly answered 88.5% of all the pages that contained a
photo for which NoOther friend matched.
4. SECURE SOCIAL AUTHENTICATION
Based on the results of our user study we proceed with
establishing guidelines for creating a secure SA mechanism
that is robust against the attacks we presented.
4.1 Tag Selection
The goal is to filter out faces that can be of use to ad-
versaries that employ face recognition software, and select
a subset of the remaining tags that have a high probabil-
ity of containing a face. We use two criteria for selecting
tags after being analyzed by face recognition software: a
high confidence level for containing a human face, and a
false recognizability flag, i.e., medium tags. While our user
study demonstrated that users are highly e↵ective even if
the friend’s face is not in the photo (i.e., di cult photos),
we do not use such photos in our SA challenges.
Even though we build our selection process using face.com,
our tag selection can be completed with any accurate face
recognition software. Once all the tags have been analyzed,
the system selects the tags that could not be recognised.
The OSN can also use several types of information to
narrow down the set from which friends are selected for the
SA challenges. This set needn’t be small, but in the order
of 200-300 friends, that have a minimum level of interaction
with the user. All friends from this set must have the same
chance of being selected, and all suggestions must be from
this set, so as not to aid the attacker by limiting the number
of suggestions that seem plausible as answers.
4.2 Tag and Photo Transformations
To defend against the image comparison attack, tagged
areas should not be identical to the areas in the original
photos, to prevent the attacker from mapping them to the
photos in the collection. Our approach blends the faces the
user has to identify with the faces on a “background” photo.
Figure 6: An example output photo, with rotation,
opacity, and perspective transformations performed.
The other faces have been blurred for privacy.
If we simply overlay the tagged areas containing the faces
onto a new photo, an adversary could still resort to the image
comparison attack. To prevent this, we perform a sequence
of transformations on the extracted areas.
Tag transformation: First, we rotate the face, a trans-
formation that can impact face detection. Second, we edit
the tag’s alpha level (a) to make it translucent and blend
it with the underlying faces (0  a  1, where 0 is fully
transparent). Thus, the tag will not contain any parts from
the photos in their initial form.
Photo transformation: Each challenge contains one
photo of N friends, with M tagged faces for each friend. We
select a background photo that contains at least N ⇤ M faces,
and overlay the existing faces with the processed tags we
created in the previous step. We then apply a perspective
transformation, which is challenging even for complex feature
or template matching techniques. According to Gauglitz et
al. [11], “perspective distortion is the hardest challenge”. The
perspective transformation we perform is variable by P, with
P denoting the ratio by which the bottom of the photo is
“compressed” from each side; e.g., for P = 3, the bottom is
compressed from both the left and right by 1/3 of the photo.
The user is presented with N menus, each one containing the
name of one of the depicted friends, along with the names
of S 1 other friends. The user is required to correctly
identify the N depicted friends. To demonstrate that familiar
faces remain recognizable after our transformations, Figure 6
shows an example output, with a = 0.6 and P = 3.2, for two
well-know politicians1
.
Prototype Implementation. We implemented a pro-
totype, which comprises of a Facebook app for the data
collection process, and a back end for the photo processing.
We implemented the back-end in Python, using SimpleCV
and OpenCV for the face detection and image processing.
To create a SA challenge, the back-end first selects N
distinct friends of the target user. For each friend, it finds
M random tags of that friend, and fetches the corresponding
photos. The tags are extracted, transformed and overlayed
on a random background photo, which is then also trans-
formed. The tag processing part randomly rotates and applies
an alpha opacity filter that ensures that none of the orig-
inal pixels of the tag are preserved. This is implemented
1
Challenge solution: Barack Obama, Vladimir Putin.

rotations CCOEFF CCORR SQDIFF time_photo
None 12.8% 11.0% 10.6% 6.61
7 (30 ) 67.8% 36.4% 43.8% 46.23
13 (15 ) 91.0% 60.0% 68.8% 87.85
19 (10 ) 95.2% 67.4% 77.8% 130.65
37 (5 ) 97.8% 75.8% 90.8% 244.68
time_rotation 2.24 2.19 2.18 -
Table 4: Attack success rate of each algorithm for
various rotation approaches, time required (sec) for
each algorithm to process one rotated version, and
the total time for all rotated versions of a photo.
with SimpleCV’s blit() function, which takes the back-
ground photo, tag image, position of the overlay and opacity
percentage as input, and returns a collage image. The rota-
tion is implemented with rotate(), the perspective trans-
formation is based on the getPerspectiveTransform() and
warpaffine() functions of OpenCV. We set a time-window
of one minute for users to solve the challenge.
4.3 Experimental Evaluation
Threat model. We assume the attacker has knowledge
of our system, and has created a collection containing all
the photos of the victim and his friends. We also assume
he can apply the same categorization to photos as we do,
and identify medium faces. Furthermore, as each tag in the
challenge has a suggestion of 6 users, the attacker will only
compare the photo to the tags of those users. In our user
study, we found that each user’s friend has ⇠12 medium
tags on average. Thus, in our experiments, for each tag, the
attacker simply has to identify which tag out of a set of 72
(12 for each suggested user) is contained in the photo, to
pass the challenge.
Image comparison attack. We employ 3 di↵erent tem-
plate matching methods: the normalized versions of the cor-
relation coe cient (CCOEFF), cross correlation (CCORR) and
squared di↵erence (SQDIFF) algorithms. To verify their ac-
curacy, we first run the attack against a set of 500 photos,
where we have overlayed a tag but have not performed any
transformations. All three algorithms correctly identify the
500 tags, requiring ⇠6.89 seconds per photo.Compared to
the simplistic pixel-comparison attack from Section 2, the
template matching algorithms identify every tag but have
two orders of magnitude processing overhead. However, the
simplistic attack cannot handle the transformations.
First, we measure the impact of rotating the tag. We
create a set of 500 photos each containing a medium tag that
has been randomly rotated within [-90 , 90 ]. In the first
attack scenario, the attacker does not perform any rotations
on the photo and simply selects the tag out of the 72 with
the highest similarity score with the photo. In the other
scenarios the attacker performs a series of 30 , 15 , 10 and
5 rotations ( in the range [-90 , 90 ] ). For each photo, the
attacker selects the best match among all the tags and all
the rotations. While the template matching algorithms can
handle certain rotations, the results of our experiment shown
in Table 4, demonstrate that e↵ectiveness is greatly increased
when multiple rotated versions of the photo are processed
before selecting the best-matching tag. CCOEFF yields the best
0
10
20
30
40
50
60
70
80
90
100
0.8
0.7
0.6
0.5
0.4
Identifiedphotos(%)
alpha value (transparency transformation)
CCOEFF
CCORR
SQDIFF
CCOEFF+rot
CCORR+rot
SQDIFF+rot
Figure 7: Attack success, against transparency and
transparency+rotation transformations.
Figure 8: Attack success against perspective (P),
and all transformations combined (RTP), for a = 0.8.
results, with a success rate of up to 97.8% when the attack
performs 5 rotations. However, the increased accuracy
comes at a cost, as it has a linear computational overhead
for the attacker. While processing the 15 rotations requires
an order of magnitude more computational e↵ort compared
to no rotations, we will use that to test the robustness of our
system (unless otherwise stated), as it is su ciently e↵ective
and 3 times faster than the attack with 5 rotations.
Next, we compare the combined impact of the rotation
and transparency transformations. We create five sets of 500
challenges, each with a di↵erent alpha level (transparency)
transformation, and run the attack without rotations. We
then create another five sets of 500 challenges each, with a
randomly rotated tag and di↵erent alpha level transforma-
tions. As can be seen in Figure 7, the CCOEFF algorithm is the
most e↵ective in all scenarios, proving to be the most e↵ec-
tive against both transparency and rotation transformations.
Nonetheless, we can see that transparency has a significant
impact, dropping the success rate of all three algorithms
to 6 10% when a = 0.4. Since such a transparency level
may prove to be di cult for users to identify the tags, we
employ more conservative transparency transformations in
the following experiments. Even with an 0.8 alpha level,
which is a mild transformation, the success rate of two of the
algorithms drops to less than 50%.
To compare the combined impact of all our transforma-
tions, we create two sets of 500 challenges each, that contain
photos with one tag; one set with tag transformations and
one with tag and photo transformations. We also place the
tags randomly in the photos, to explore the significance of
the “background” photo. We maintain a constant 0.8 al-
pha level (which is a very conservative transformation and
remains easily identifiable to humans), and experiment by

0
3
6
9
12
15
18
6
5
4
3.2
2.7
Identifiedphotos(%)
Perspective transformation
CCOEFF-a0.8
CCORR-a0.8
SQDIFF-a0.8
CCOEFF-a0.7
CCORR-a0.7
SQDIFF-a0.7
CCOEFF-a0.6
CCORR-a0.6
SQDIFF-a0.6
Figure 9: Attack success rate against di↵erent per-
spective (P) and alpha level transformations.
varying the perspective transformation (P = 2.7, .., 6). Fig-
ure 8 shows that the perspective transformation (P) has the
most significant impact, but the tag transformations (RTP)
also contribute to the overall robustness, especially when the
perspective transformation is not as extreme, i.e., for larger
values of P. SQDIFF proves to be the most e↵ective method
against the perspective transformation, while CCOEFF is the
least e↵ective despite being the most robust against trans-
parency and rotation transformations. We achieve our best
results when P = 3.2, 2.7, with the highest scoring method
(SQDIFF) identifying only ⇠7.4% of the photos that have
undergone all transformations, compared to the 98.4%-98.8%
success rate of our simplistic attacker (see Section 2), or
the 100% rate of the template matching attack against non-
transformed photos. Depending on the placement of the tag
in the photo and the rotation e↵ect, the perspective trans-
formation has a combined impact that cannot be handled by
the attacker even when conducting multiple rotations.
To further explore the combined impact of the transforma-
tions, in Figure 9 we show the percentage of identified photos
for each algorithm with varying alpha level and perspective
transformations. Our results clearly show a correlation be-
tween the identified tags and the levels of transformation. As
the alpha and P values decrease (i.e., the tag becomes more
transparent and the photo’s perspective is more distorted)
the e↵ectiveness of the attack decreases by up to a factor of
two for SQDIFF and by up to a factor of five for CCOEFF. Apart
from the impact on the success rate, the transformations
require a significant increase of processing e↵ort from the
attacker. Attempting to match the tags to the transformed
photo requires ⇠87.8 as opposed to ⇠0.06 required for the
simplistic attack against the non-transformed photos.
Surprisingly, we found that for a perspective transforma-
tion of 2.7 we didn’t see a significant decrease compared
to 3.2 and CCORR actually scored higher. As such, we set
P = 3.2 for our next experiments as it is less distorting and
also yields slightly better results. Furthermore, we manually
inspected the identified tags and found them to be cases
where they had been placed randomly on a position of the
background photo with almost “no noise” (e.g., on a clear sky,
wall, etc.). Thus, we should strongly enforce a restriction of
placing the tags on faces in the background photos, which
will further decrease success rates.
We measure the e↵ect of increasing the number of tags per
challenge. We create three sets of 1,000 photos that contain
two tags (N = 2), with P = 3.2 and varying alpha levels.
The attacker now has to compare each photo to 144 tags (72
Tags alpha CCOEFF CCORR SQDIFF Time (sec)
2
0.8 0.0% 2.1% 1.8% 397.7
0.7 0.0% 1.9% 1.6% 400.6
0.6 0.2% 1.5% 1.0% 401.5
3
0.8 0.0% 0.0% 0.0% 663.6
0.7 0.0% 0.0% 0.4% 675.0
0.6 0.0% 0.4% 0.0% 695.9
Table 5: Attack success with 5 rotations against
challenges with 2 tags, for P = 3.2.
Detected Faces Before
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
DetectedFacesAfter
0
1
2
3
4
5
6
7
8
9
10
11
12
13
886 2381 149 50 33 10 5 2 3
543 1 2398 153 83 33 22 9 7 2
26 2 1 1168 141 50 22 13 14 6 2
2 1 2 112 40 26 25 11 9 7 4
2 2 3 15 14 13 6 8 8 4
5 2 1 1 2 22 4 1 2 1 3
1 1 1 11 2 3
1 1 11
1 1 1 1 11
1
1 1 1 1 1
1 1
1
Figure 10: Faces detected before and after rota-
tion+transparency. Point labels correspond to the
number of photos. The line shows the X = Y axis.
per depicted friend), and correctly identify both tags to pass
the challenge. To calculate the maximum success rate the
attacker can achieve, the attack conducts 5 rotations which
produce the best results, even though the processing time
is unrealistic. Table 5 shows the results. With a 0.6 alpha
level, CCOEFF fails to pass any challenges, CCORR passes 15
and SQDIFF passes 10 challenges, while processing a photo
requires ⇠ 401.5 sec. We create three sets of 250 photos
with 3 tags, where one algorithm fails and the rest solve a
single challenge, requiring over ⇠ 663.6 seconds. Thus, we
reduce the attacker’s success rate to 0.4%, while requiring
four orders of magnitude more processing e↵ort on his part.
We also explored the possibility of combining the results of
the 3 methods for achieving better success rates, i.e., compare
the output of each method and select the photo that receives
the highest confidence out of the 3. This, however, is infea-
sible because in our experiments the “confidence” returned
by CCORR is always higher than the one by CCOEFF, even
when CCORR’s identification is wrong. Also, SQDIFF returns a
di↵erential result which is not comparable to the other two.
Face Recognition attack. We also evaluate the robust-
ness of our approach against face detection and, by extension,
recognition. To explore how our tag selection and transforma-
tion process impacts face detection, we calculate the number
of faces detected in 3,487 “background” photos before and
after we transform the tags and superimpose them on the
background photo (no perspective transformation performed).
We first detect the faces in the photo, then superimpose a
transformed tag over every face and, finally, execute the face
detection process again. We perform a conservative trans-
parency transformation with an alpha level of a = 0.8 that
can easily be solved by users. Figure 10 shows the detected
faces before and after, and the label of each point indicates
the number of photos with that (before, after) tuple. The

0
3
6
9
12
15
0.8
0.7
0.6
0.5
0.4
DetectedFaces(%)
alpha value (transparency transformation)
Transparency
Transparency+Perspective
Figure 11: Faces detected for various alpha levels,
against transparency and transparency+perspective
transformations.
red line denotes the X = Y axis, upon which are the cases
where the same number of faces are detected. Interestingly,
even though the photos now contain double the number of
faces, an extra face is detected only in 47 (1.3%) cases (points
over the red line). Everything below the red line, indicates
improvement as less faces are detected. Due to our tag trans-
formations, no faces are detected in 43.6% of the photos,
which significantly impacts face recognition, as faces have to
be detected before compared to facial models.
While the rotation transformation increases the processing
e↵ort of the attacker, it cannot hinder an attacker as a stand-
alone transformation, as the attacker can perform rotations
upon the crafted photo to increase detection rates. Thus,
we want to explore the combined impact of our two other
transformations. To do so, for each experiment we create
two versions of 250 transformed photos with one medium tag
each. In the first version, we only apply the transparency
transformation, and in the second both the transparency
and perspective transformation. We then manually remove
any photos with tags that do not contain a human face,
and use our face detection method to see how many faces
are detected in the remaining tags. We test various alpha
levels of transparency, with a constant value of P = 3.2
for the perspective transformation, as it had the best e↵ect
in our previous experiments. We present our results in
Figure 11. While the transparency transformation has a
very significant impact on the attack, by combining it with
perspective transformation, we are able to completely hinder
face detection and, thus, recognition in all the photos, for
a  0.6. Again, we found that any faces that were detected
had been placed on a “clear” background.
Security through distortion. Overall, our experiments
demonstrate the robustness of our transformations against
pattern matching and face recognition attacks. Even under
an extreme threat model where the attacker has collected
every single photo of the users and has knowledge of our
approach, we are able to completely deter one class of attacks
and decrease the success rate of the other by over 99%, while
incurring a massive overhead to the required processing e↵ort.
4.4 Usability Evaluation
To evaluate the usability of our approach we conduct a
preliminary study using challenges with tags of famous people.
While users might be more familiar with certain famous
people than with some of their contacts, they may not be
familiar with others as well (or know their name). The goal
Departmental AMT
Initial Normalized Success
Gender NoHelp Help NoHelp Help 94.38%
Female 75.0% 95.8% 76.4% 97.8%
Male 73.9% 87.5% 78.5% 92.6% Time
Total 74.2% 89.7% 77.9% 94.0% 7.93s
Table 6: Usability Evaluation. Initial and normal-
ized success rates per gender, before and after sug-
gestions are presented (Departmental). Success rate
and seconds per challenge (AMT).
of this study is to measure the impact of the transformations
on the ability of users to identify the depicted people. As
such, we have selected fairly recognizable individuals.
We selected photos of them wearing sunglasses, with their
face at an angle, taken under normal conditions and not
edited. We used the values that we found in our experimen-
tal evaluation to be secure, but not to extreme so as to impede
users’ ability for identification. Specifically, we used all the
possible combinations of: N = {1, 2}, a = {0.6, 0.7, 0.8}, P =
{2.7, 3.2}. We conduct two separate studies, one with par-
ticipants from our department, and one with Amazon Turk
workers, each case allowing us to measure di↵erent aspects of
our system; the impact of suggestions and the time required
to solve the challenges.
Departmental. We recruited 30 participants from our
department (students and sta↵), 8 of which were female and
22 were male, with ages ranging from 19 to 37. The same
set of 12 challenges (one for every possible combination) was
presented to all participants, to identify whether a specific
transformation or person could not be identified by multiple
users. Users were first shown a photo, and if they were able
to provide the correct answer they would move on to the
next. If not, they were given 6 suggestions for each tag,
which allowed us to measure the impact of the suggestions.
After the initial results, we also calculated a normalized rate
where we filtered out challenges where the participant did
not know the depicted person at all, as our goal is to measure
the impact of the transformations alone.
As shown in Table 6, users solved 89.7% of the challenges,
which increases to 94% if normalized. Surprisingly, users
immediately identified over 75% of the challenges without
help. Suggestions do o↵er a significant boost, with 14.1% and
21.4% for male and female participants respectively. There
was no strong correlation between the values of transforma-
tions and the success rates. On the contrary, less transformed
photos had higher scores in several cases. We found, however,
that the face was an important factor as there were a few tags
that users could not identify even without transformations, as
they were barely familiar with certain individuals. Nonethe-
less, suggestions are an important aspect of the system as
they help users solve the challenges.
AMT study. We recruited AMT workers (at least 95%
approved, correct HITs and 1000 HITs completed) and pre-
sented them with two batches of 72 distinct challenges gen-
erated automatically by our system. The first batch had
N = 1 celebrity face and the second one had N = 2. For each
batch we included all the possible combinations of a and P as
described above, and 6 suggested names (of which 1 was the
correct answer). Overall, 49 workers solved 1,556 challenges,

with at least 20 workers solving each challenge. As Table 6
shows, AMT workers confirmed the previous results (on a
larger scale) with a 94.38% success rate (not normalized),
taking 7.93 seconds on average (7.19 standard deviation).
Conclusions on Usability. A crucial metric for the ap-
plicability of such a mechanism is the ability of users to solve
the challenges. The results of the user study demonstrate
the usability of our system, as users solved ⇠94% of the
challenges. Bursztein et al. [8] reported an 84–87% success
rate for image CAPTCHAS, which is lower than our results.
For the easiest category of text CAPTCHAs results were
slightly higher than ours, with 95–98%. Furthermore, they
reported an average completion time of ⇠ 9.8 seconds for the
image CAPTCHAs (⇠ 8.9 when solved correctly), which is
almost two seconds slower than ours.
We believe that our initial results demonstrate that our
system provides a viable solution for securing the login pro-
cess against adversaries that have stolen user credentials.
Nonetheless, we plan on conducting an extensive user study
with challenges that depict actual contacts of the users, to
fully evaluate the usability of our approach.
5. SOCIAL AUTHENTICATION SERVICE
We discuss how an OSN can deploy our secure SA as
a service for other websites. By adding a simple piece of
code, services can access an API and fetch challenges specifi-
cally crafted for each individual user. This can be adopted
as a user-gnostic CAPTCHA, or as an additional security
mechanism.
To outline the benefit of employing such a service in ad-
dition to a traditional two-factor authentication scheme, we
describe the following scenario. An attacker steals a user’s
smartphone, which contains the credentials to an e-banking
service and is also the device that receives the security token
(via SMS or a token-generation app). Normally, the attacker
will be able to complete any transaction as he possesses
both tokens needed to pass the two-factor authentication.
Similarly, attacks in the wild, have passed two-factor authen-
tication by tricking the user to install a malicious app (e.g.,
the Eurograbber malware [2]). However, if the e-banking
service employs this service for additional security, attackers
that don’t belong to the victim’s social circle will fail to
complete any transaction. Even if the device has an exist-
ing session with the OSN, they will not be able to pass the
challenges (see outsourcing attacks below).
Dataset. Another important feature of our system, is that
it builds upon a dataset of photos and tag information that
is readily available to the OSN. Thus, it doesn’t face the
challenge of creating a correct and extensive dataset as other
image-based CAPTCHA schemes do [15, 26].
Someone could argue that SA is too restricted in terms of
the number of challenges that can be created, in comparison
to text-based CAPTCHA schemes than can create infinite
combinations of numbers and letters. However, such an
argument is far from true for our approach. In our study we
found that users have an average of 347 friends each with
approximately 12 medium tags, resulting in 4,164 suitable
tags per user. These can produce over 8 million di↵erent
permutations of 2 tags, and 1.2e+10 for 3 tags, which is more
than enough for a single user profile. Also, the huge number
of photos that can be used as “backgrounds” increases the
number of possible challenges even more.
Privacy. A user-gnostic CAPTCHA service may raise
privacy concerns, as the OSN acquires information regarding
websites visited by users. However, this information is also
acquired through the“social plugins”or single sign-on services
o↵ered by many popular OSNs. These services have been
widely adopted, and [17] reports that over 35% of the top
10,000 Alexa websites include the “Like” button.
Security Properties. We discuss the e↵ectiveness of our
approach against typical CAPTCHA-breaking attacks.
Guessing Attacks. Our scheme allows automated bots to
pass the challenge with a probability of 1/SN
, where N is the
number of friends depicted and S the number of suggestions
per friend. The threshold adopted by previous work [28] is
that bots should not have a success rate higher than 0.6%.
By selecting 3 friends and providing 6 suggestions per friend,
we are able to achieve an even lower probability of 0.46%.
Furthermore, our system provides an extra level of security.
As each CAPTCHA is created for a specific user, it is trivial
to identify automated attacks that try to guess the answer.
A legitimate user can request a new challenge when not able
to identify the friends, without providing an answer, until
presented with one he feels confident about answering. An
automated script trying to guess the answer will continuously
provide wrong answers until eventually guessing one.
Outsourcing/Laundering Attacks. Existing approaches cre-
ate challenges that are user-agnostic, i.e., are not created for
a specific user. They are built upon the notion that virtually
any human should have a high probability of successfully car-
rying out the tasks required to pass the challenge. However,
these approaches are susceptible to “laundering” attacks [19]
where adversaries relay the challenge to CAPTCHA-breaking
services with human solvers that provide the correct an-
swer, and ‘smuggling” attacks [9] that trick users into solving
CAPTCHAs by injecting them in interactions with benign
sites. Our approach is robust against such attacks, as chal-
lenges are user-gnostic: they cannot be outsourced to others,
as they wouldn’t be familiar with the user’s friends. Solving
them would require the worker to first familiarize with the
friends’ faces. This, of course, is impractical as it would
require too much time to pass a single CAPTCHA challenge
(might not even be possible within the allowed time window).
6. LIMITATIONS AND FUTURE WORK
The attacker could create a collection of processed tags and
compare those to the presented SA challenge. However, vari-
ous characteristics of our approach render it robust against
such a scenario. First, the completely random background
photo, which blends in with the tags, introduces significant
noise which can’t be predicted by the attacker. Second, the
placement of the tag on a photo significantly a↵ects the result
of its perspective transformation. Finally, as the transforma-
tions’ values can be selected from a range, the attacker would
have to create a massive collection of processed tags with
various combinations of transformations and backgrounds.
Even then, identifying the tag might not be feasible.
The attacker could attempt to reverse the perspective
transformation. However, that requires knowledge of the
exact value of the transformation, which can be randomly
selected from a range of values. Furthermore, various types of
perspective transformations exist (which we plan to explore),
and our system could randomly select one for each challenge.
Even in the implausible scenario where an attacker could
guess the correct combination of algorithm and exact value,

the transparency and rotation transformations provide a
very strong defence against the face detection attack and can
reduce the success of the image comparison attack by over
80% for two of the algorithms.
7. RELATED WORK
The first to analyze SA and discuss potential vulnerabilities
were Kim et al. [16], who presented a formal quantification
of the risks that threaten SA. A key observation is that
tightly connected communities exhibit higher risks. In [20]
we demonstrated the feasibility of attacks against SA: a casual
attacker with access only to publicly available information
would pass 22% of the challenges and significantly improve
the chances for 56%; a determined attacker, with access to
social-circle information, would pass 100% of the challenges,
when training classifiers with 120 faces per friend.
Bursztein et al. [8] conducted an extensive study to measure
the success rate of users when presented with CAPTCHAs
from various services. An important observation the authors
make is that the di culty of CAPTCHAs is often very high,
resulting in their solution being a troublesome process for
users. A significant aspect of our approach is that it is
user-friendly, as users are required to identify their friends.
Previous work [12, 22] has explored the use of human faces
for creating CAPTCHAs. Goswami et al. [12] built upon
the failure factors of face detection algorithms to propose a
CAPTCHA that uses distorted human and cartoon faces as
a challenge, and users must identify all human faces without
making any mistakes. The pool of faces is compiled from
publicly available databases and distortions are applied, such
as adding stripes to cover key face characteristics, rotating
the image and overlaying the actual face over a background
of random noise. While this approach may be robust against
face detection, it can be bypassed with image comparison
algorithms (like template matching), that map the visible
parts of the faces to the photos within the public database.
Previous work (e.g., [26]) has attempted to defend against
face recognition attacks, based on semantic properties of the
photo’s content. This presents a significant drawback typical
of various image-based CAPTCHA systems; the creation of
the dataset labelled with the semantic description cannot
be automated. On the contrary, SA takes advantage of an
enormous, ever-expanding, labelled dataset created by users.
8. CONCLUSIONS
In this paper we revisited the concept of social authentica-
tion, and proposed a novel approach that retains its usability
while being robust against attacks that utilize image anal-
ysis techniques. The key concept is to filter out faces that
can be identified by face recognition software, and craft the
challenge photos in a manner that obfuscates any areas with
sensitive information from the initial photos. We conducted
a measurement to explore the ability of users to identify their
friends in photos taken under realistic conditions. Our re-
sults demonstrated that users are very e↵ective at identifying
their friends even when their face is not clearly visible or
present at all. Based on our study results and a series of
observations, we implemented a prototype and evaluated it
against the face recognition and image comparison attacks.
We also verified the usability of our system through a user
study. Finally, we discussed the benefits of employing such a
system as a user-gnostic CAPTCHA service, or an additional
security mechanism to two-factor authentication.
Acknowledgements
We thank the anonymous reviewers for their valuable com-
ments. This work was supported in part by the FP7 project
SysSec funded by the EU Commission under grant agreement
no 257007, the FP7 Marie-Curie ITN under grant agreement
no 316808, the ForToo Project of the Directorate General
Home A↵airs, and by the MIUR under the FIRB2013 FACE
grant. This work was also supported by the NSF Grant
CNS-13-18415. Author Keromytis was also supported by
(while working at) the NSF during the conduct of his work.
Any opinions, fundings, conclusions, or recommendations ex-
pressed herein are those of the authors, and do not necessarily
reflect those of the US Government or the NSF.
References
[1] Amazon Mechanical Turk.
https://www.mturk.com/mturk/.
[2] Eurograbber.
https://www.checkpoint.com/products/downloads/
whitepapers/Eurograbber_White_Paper.pdf.
[3] Facebook Acquires Face.com. http://mashable.com/
2012/06/18/facebook-acquires-face-com/.
[4] Facebook Introducing Login Approvals.
https://www.facebook.com/note.php?note_id=
10150172618258920.
[5] Google 2-step.
http://www.google.com/landing/2step/.
[6] L. Bilge, T. Strufe, D. Balzarotti, and E. Kirda. All
your contacts are belong to us: automated identity
theft attacks on social networks. In Proceedings of the
18th International Conference on World Wide Web.
ACM, 2009.
[7] Y. Boshmaf, I. Muslukhov, K. Beznosov, and
M. Ripeanu. The socialbot network: when bots
socialize for fame and money. In Proceedings of the
Annual Computer Security Applications Conference.
ACM, 2011.
[8] E. Bursztein, S. Bethard, C. Fabry, J. C. Mitchell, and
D. Jurafsky. How good are humans at solving
CAPTCHAs? A large scale evaluation. In Proceedings
of the 2010 IEEE Symposium on Security and Privacy.
IEEE, 2010.
[9] M. Egele, L. Bilge, E. Kirda, and C. Kruegel. Captcha
smuggling: Hijacking web browsing sessions to create
captcha farms. In Proceedings of the 2010 ACM
Symposium on Applied Computing, SAC ’10’, pages
1865–1870. ACM, 2010.
[10] M. Egele, G. Stringhini, C. Kruegel, and G. Vigna.
COMPA: Detecting Compromised Accounts on Social
Networks. In ISOC Network and Distributed System
Security Symposium (NDSS), 2013.

[11] S. Gauglitz, T. Höllerer, and M. Turk. Evaluation of
interest point detectors and feature descriptors for
visual tracking. Int. J. Computer Vision,
94(3):335–360, 2011.
[12] G. Goswami, B. M. Powell, M. Vatsa, R. Singh, and
A. Noore. FaceDCAPTCHA: Face detection based
color image CAPTCHA. In Future Generation
Computer Systems (September 2012).
[13] J. Huang, Y. Xie, F. Yu, Q. Ke, M. Abadi, E. Gillum,
and Z. M. Mao. Socialwatch: detection of online
service abuse via large-scale social graphs. In
Proceedings of the 8th ACM SIGSAC symposium on
Information, computer and communications security,
ASIA CCS ’13’, 2013.
[14] D. Irani, M. Balduzzi, D. Balzarotti, E. Kirda, and
C. Pu. Reverse social engineering attacks in online
social networks. In Proceedings of the 8th international
conference on Detection of intrusions and malware, and
vulnerability assessment, DIMVA’11, 2011.
[15] E. Jeremy, J. R. Douceur, J. Howell, and J. Sault.
Asirra: a CAPTCHA that exploits interest-aligned
manual image categorization. In Proceedings of the
14th ACM conference on Computer and
communications security (CCS). ACM, 2007.
[16] H. Kim, J. Tang, and R. Anderson. Social
authentication: harder than it looks. In Proceedings of
the 2012 Financial Cryptography and Data Security
conference. Springer.
[17] G. Kontaxis, M. Polychronakis, A. D. Keromytis, and
E. P. Markatos. Privacy-preserving social plugins. In
Proceedings of the 21st USENIX conference on Security
symposium, Security’12. USENIX Association.
[18] Y. Liu, K. P. Gummadi, B. Krishnamurthy, and
A. Mislove. Analyzing facebook privacy settings: User
expectations vs. reality. In Proceedings of the 2011
ACM SIGCOMM Conference on Internet Measurement
Conference, IMC ’11’. ACM, 2011.
[19] M. Motoyama, K. Levchenko, C. Kanich, D. McCoy,
G. M. Voelker, and S. Savage. Re: Captchas:
understanding captcha-solving services in an economic
context. In Proceedings of the 19th USENIX conference
on Security, USENIX Security’10. USENIX Association,
2010.
[20] I. Polakis, M. Lancini, G. Kontaxis, F. Maggi,
S. Ioannidis, A. Keromytis, and S. Zanero. All your
face are belong to us: Breaking facebook’s social
authentication. In Proceedings of the 28th Annual
Computer Security Applications Conference, ACSAC
’12’. ACM, 2012.
[21] A. Rice. Facebook - A Continued Commitment to
Security, Jan 2011. http:
//www.facebook.com/blog.php?post=486790652130.
[22] Y. Rui and Z. Liu. Artifacial: Automated reverse
turing test using facial features. In In Multimedia,
pages 295–298. ACM Press, 2003.
[23] A. Shulman. The underground credentials market.
Computer Fraud & Security, 2010(3):5–8, March 2010.
[24] Y. Taigman and L. Wolf. Leveraging billions of faces to
overcome performance barriers in unconstrained face
recognition. CoRR, abs/1108.1122, 2011.
[25] B. E. Ur and V. Ganapathy. Evaluating attack
amplification in online social networks. In Proceedings
of the 2009 Web 2.0 Security and Privacy Workshop.
[26] S. Vikram, Y. Fan, and G. Gu. SEMAGE: A New
Image-based Two-Factor CAPTCHA. In Proceedings of
2011 Annual Computer Security Applications
Conference (ACSAC’11), December 2011.
[27] S. Yardi, N. Feamster, and A. Bruckman. Photo-based
authentication using social networks. In Proceedings of
the first workshop on Online social networks, WOSN
’08’. ACM, 2008.
[28] B. B. Zhu, J. Yan, Q. Li, C. Yang, J. Liu, N. Xu,
M. Yi, and K. Cai. Attacks and design of image
recognition captchas. In Proceedings of the 17th ACM
conference on Computer and communications security,
CCS ’10’. ACM, 2010.
APPENDIX
Appendix A
Below is the questionnaire that users were requested to fill
in for each photo presented to them during the Social
Authentication challenges. Assuming the depicted friend’s
name is X, the user will be presented with the following
questions.
1. Type of photo?
• Portrait
• Landscape
• Objects
• Text
• Animals
• Art
2. Where is X’s Face?
• Within the tag and is clearly visible.
• Outside the tag and is clearly visible.
• Within the tag, but not clearly visible.
• Outside the tag, but not clearly visible.
• Not in the photo at all.
3. Are there other faces in the photo?
• There are other people’s faces both outside and inside the
tag.
• There’s someone else’s face within the tag.
• There’s someone else’s face outside of the tag.
• There are no other faces in this photo.
• There are no faces in this photo.
4. Why was this photo useful for identifying X?
• I remember seeing this photo from X.
• The content of the photo is relevant to X.
• None of the other suggested friends matched.
• This photo was not useful.
• X is in the photo.

Faces in the Distorting Mirror: Revisiting Photo-based Social Authentication

More Related Content

What's hot (17)

Similar to Faces in the Distorting Mirror: Revisiting Photo-based Social Authentication (20)

More from FACE (9)

Recently uploaded (20)

Faces in the Distorting Mirror: Revisiting Photo-based Social Authentication