ppt - Department of Computer and Information Sciences

MobiFish: A Lightweight Anti-Phishing
Scheme for Mobile Phones
1
Longfei Wu, Xiaojiang Du, and Jie Wu
Department of Computer and Information Sciences
Temple University, Philadelphia, PA, 19122, USA
1/12/2017
Presenter: Dr. Xiaojiang (James) Du
Phishing Attacks
2

Phishing attacks aim to steal private information
such as usernames, passwords, and credit card
details by impersonating a legitimate entity.

Although security researchers have proposed many
anti-phishing schemes, phishing attacks’ threat has
not been well mitigated:



Phishing sites expire and revive rapidly (Avg. 4.5 days).
Attackers keep improving their techniques to circumvent
existing anti-phishing tools.
Mobile users are accustomed to being requested and
providing credentials without checking the website.
1/12/2017
Phishing Attacks
3

Most targeted Industry Sectors
1/12/2017
Phishing Attacks Cont.
4

Almost all phishing attacks on PC are in the form of bogus
websites. Current browsers on PC are embedded with antiphishing tools that can achieve a detection rate of over 90%.

However, during the adaptation to hardware-constrained
mobile platforms, browsers abandoned or truncated many
features and useful functions (like anti-phishing).
Open the same phishing site with Chrome on PC and Chrome for Android
Mobile Phishing Attacks
5

Mobile Web Phishing



Mobile phishing is an emerging threat targeting at mobile users of
financial institutions, online shopping and social networking
companies.
Mobile App Phishing

Some attackers develop fake applications (Apps) or repackage
legitimate Apps, then upload these phishing Apps to unofficial app
markets.

It is harder to detect Phishing Apps than Phishing on mobile web
pages. (Information can be retrieved from Html source code in
webpages).
The trend of launching phishing attacks on mobile devices can
be attributed to hardware limitations such as small screen
size, and the inconvenience of user input and application
1/12/2017
switching.
Existing Phishing Detection Schemes
6

Current web phishing detection schemes can be roughly
divided into two categories: heuristics-based schemes
and blacklist-based schemes.


Blacklist-based schemes can only detect phishing sites that are
in the blacklist but can not detect zero-day phishing attacks.
Heuristics-based schemes largely depend on features extracted
from URL and HTML source code, and other techniques like
machine learning are used to determine the validity.


However, we find that features extracted from HTML source code could
be inaccurate and phishing sites can circumvent those heuristics.
There is no off-the-shelf tool to detect phishing Apps on
mobile platform.
1/12/2017
Our Solutions and Contributions
7

We propose MobiFish, a novel automated lightweight antiphishing scheme for mobile phones.


It is able to defend against both phishing webpages and Apps.
Find the weakness of previous heuristics-based security
schemes for webpage phishing, and develop a lightweight
solution that utilizes optical character recognition (OCR)

without reliance on HTML source code, search engine or
machine learning techniques.

Implement MobiFish on Google Nexus 4 smartphone
running Android 4.2 operating system.

Evaluate MobiFish with 100 phishing URLs and
corresponding legitimate URLs,

as well as “Facebook” phishing Apps.
1/12/2017
Mobile Webpage Phishing Attacks
8

Mobile user interface increases the vulnerability to
mobile phishing attacks.



Due to the small display size of phone screens, most
mobile browsers have to remove the status bar and hide
the URL bar once the web page finishes loading.
Even during the loading process, long URLs are truncated
to fit the browser frame.
Since the ability to read and verify URLs is crucial in
detecting phishing attacks, partial URL or even URL
displayed with partial domain name would certainly
increase the risk of being spoofed by phishing attacks.
1/12/2017
Mobile Application Phishing Attacks
9

Application-oriented phishing attacks can be categorized
into two types based on the way they launch:

Some phishing apps attempt to hijack existing legitimate targets.



Another type of phishing apps directly appears as the target app.


They keep performing task polling, and launch themselves as long as
they detect the launch of target apps.
As the result, the fake login interface covers on top of the real one,
and the phishing app pretends to be the target app.
This may occur when user downloads fake apps from unofficial app
markets.
The mobile App phishing attack ends with transmission
of credentials to the attacker.

Hence, blocking the transmission can effectively defend the
attack.
1/12/2017
Overview of MobiFish Scheme
10



Phishing attackers apply fancy tactics to direct
victims to their phishing sites or applications, which
masquerade as trustworthy entities.
The key to solve phishing problem is to find the
discrepancy between the identity it claims and the
actual identity.
MobiFish consists of two independent components
designed for mobile webpages and mobile
applications

WebFish and AppFish.
1/12/2017
Design of WebFish
11

We find that information extracted from HTML source
code may not reflect the webpage displayed to users,





since attackers can add texts, images and links into HTML
source code while making any “undesirable” content invisible,
by simply changing their size or covering them with other
images.
Hence, features like word frequency, brand name and
company logo could be easily manipulated.
The claimed identity should be extracted from the
screen presented to a user.
The actual identity can be obtained from the web
address (or network connection).
1/12/2017
Identity Extraction
12

The claimed identity is extracted from a screenshot.

Most login interfaces of legitimate mobile sites and apps are very
simple. The entire login page or the majority of page can be captured
in one screenshot.

To obtain claimed identity from a screenshot, OCR technique is
utilized to convert image into text.
We use Tesseract, one of the most accurate open source OCR engines.


The actual identity is obtained from the web address.

Most enterprises use brand name as the second-level domain name
(SLD) of their official websites.

In cases that brand names are not exactly the same as SLD (e.g.
brand name “AT&T” and SLD “att”), we build a whitelist that records
common pairs of inconsistent brand name and SLD.

brand name “AT&T” is directly mapped to SLD “att”, and vice versa.
1/12/2017
Identity Extraction Cont.
13

OCR Experiments


Our testing uses a Thinkpad T420 laptop (2.40GHz, 4GB RAM)
with pixel density of 131 dpi and a Google Nexus 4 smartphone
(1.5GHz, 2GB RAM) with 320 dpi pixel density.
We open the Ebay mobile login page in both mobile and PC
browsers, each captures a screenshot. Then, Tesseract is used
to extract text from phone screenshot while Microsoft Office
Document Imaging (MODI) is used for the screenshot on PC.

Tesseract


MODI
Tesseract only takes 1.6 seconds while MODI uses 4.5 seconds.
1/12/2017
Design of WebFish Cont.
14

Finally, WebFish compare the claimed identity with
the actual identity.
1/12/2017
Design of WebFish Cont.
15

The key idea of WebFish to detect a phishing URL is that the SLD is
not among the text extracted from the screenshot of the login page.

As far as we know, no phishing site uses common terms in login
pages like “sign”, “username”, “password” or “welcome” as SLD.

It is not likely for well constructed and maintained legitimate web
pages to have strange words.

If the actual domain name of a phishing site appears in the login
page of fake websites, users can easily spot it and check the URL
to verify the validity of the webpage.

If the attacker includes the phishing domain name in the screen in a
tiny font size, then OCR is not able to recognize it either and
WebFish will still mark it as a phishing site.
1/12/2017
Design of AppFish
16

AppFish maintains a database called suspicious app set
(SAS), which contains profiles of untrusted apps


including user ID (Uid), launching time and screenshot text.
These apps should be:



Specified for one company. This is to ensure that the app only
connects to the company’s official sites or affiliated (partners)
servers.
The domain name of collaborators are pre-checked and added to
the SAS profile in advance. (e.g. Facebook and its content
delivery networks)
Have user login. There are lots of apps that do not need users to
login, in which App phishing attacks would not happen at all. (e.g.
apps for news, games, music or map)
1/12/2017
Design of AppFish Cont.
17

Phishing apps are not able to load valid following pages.
Users will suspect their validity in a short time.


AppFish monitors the possible paths that allow a phishing
app to transmit data to outside,


Hence, a phishing app can only send out user credentials during a
short period (denoted as T) after user clicks the phishing page.
which include socket, HttpGet/HttpPost, SMS, email (email is based
on socket), etc.
AppFish rules:


The SLD name of the Http connection destination has to be in
the text or affiliated domain names stored in SAS profile.
Socket and SMS function could be blocked for a period of time,
which should be long enough for user to notice (and uninstall)
the phishing app.
1/12/2017
Design of AppFish Cont.
18

The AppFish defense scheme works in two phases:
launching phase and authentication phase.
1/12/2017
Performance Evaluation
19

We implement MobiFish on a Nexus 4 smartphone. We
modify the source code of Android 4.2 system so that it is
able to support MobiFish.

Experiments with WebFish

We randomly pick up 100 phishing URLs from PhishTank.com.

Most of them are highly similar to their legitimate counterparts.

The input forms in phishing login pages are often surrounded by
brand names or company logos as the legitimate login pages.

When loading a large conventional web page, mobile browsers
often display the area that contains the input form instead of
displaying an overview of the entire web page.
1/12/2017
Performance Evaluation Cont.
20

WebFish is able to detect all the phishing webpages and
achieves 100% verification rate of legitimate URLs
1/12/2017
Performance Evaluation Cont.
21

Experiments with AppFish

There are only a few reported phishing apps and none of them is
available online.

To test the effectiveness of AppFish, we develop two sample
phishing apps: one can hijack real Facebook app and the other
appears as “Facebook”.

After user clicks the “Log in” button,
the fake apps send the credentials
to our server by HttpGet, HttpPost,
socket, SMS, and email, respectively.

AppFish can block all the connections
and warn users about the phishing
attempts.
1/12/2017
Conclusion
22

We proposed MobiFish, a novel lightweight mobile phishing
defense scheme.

MobiFish uses OCR, which can accurately extract text from
the screenshot of mobile login interface so that the claimed
identity is obtained. Mobile phones have higher dpi than PC.

Compared to existing OCR-based anti-phishing schemes
(designed for PC only), Mobifish is lightweight and it works
without using external search engines or machine learning
algorithms.
We implemented MobiFish on a Google Nexus 4
smartphone, and conduct experiments, which show that
MobiFish and AppFish can effectively detect and defend
against mobile phishing attacks.
1/12/2017

Thank You!
23
Prof. Xiaojiang (James) Du
Dept. of Computer and Information Sciences
Temple University
Philadelphia, PA, 19122, USA
Email: [email protected]
Web: www.cis.temple.edu/~xjdu
1/12/2017