Using Text Mining to Infer the Purpose of Permission Use in Mobile Apps

Using Text Mining to Infer the Purpose of
Permission Use in Mobile Apps
Haoyu Wang, Peking University
Jason Hong, Carnegie Mellon University
Yao Guo, Peking University
UbiComp 2015, Osaka, Japan

+++++++++++++++++++++++
 tempMobile apps can access a wealth of
sensitive data and sensors

Android Permission
 Android currently requires developers to declare
what permissions an app uses
 e.g., to access precise location need to request
“ACCESS_FINE_LOCATION” permission
 Mobile users could also set permission preference
for each app
 AppOps: Android 4.3
 Third-party permission manager

What are your apps really doing?
 An app could use a permission for multiple
purposes
Location for:
advertising, map, nearby
searching, and geotagging
POCO food camera

What are your app really doing?
 Many apps have unusual permissions
Location Data
Unique Device ID
Brightest Flashlight
Location Data
Unique Device ID
Network State
Holy Bible

However, Android offers no mechanisms
to specify the purpose of how the
permission (sensitive data) will be used

It’s important to know the purpose
 Knowing the purpose of permission use could:
Help respect to privacy:
offering end-users more
insights as to why an app is
using a specific sensitive data. Fine-grained access control

How to infer the purpose of permission use?

Related Work(1)
 Inferring permission use from app description
 WHYPER[USENIX Security ’13] and AutoCog[CCS ’15]
propose to use NLP techniques to infer permission use from app
descriptions
 Limitation
 However, for more than 90% of apps, it’s impossible to
understand why permissions are used based solely on
app descriptions

Related Work(2)
 Inferring purpose by analyzing third-party
libraries
 Lin et al. [UbiComp’12] proposed to infer the
purpose of a permission request by analyzing third-
party libraries used in an app
 Limitation
 Most permission uses occur in custom code
 e.g., for apps that use the contacts permission, more
than 71.2% of them use it in their custom code
(privacygrade.org)

How to infer the purpose of a permission
use in custom-written code?

Our Approach
 Focus on inferring the purpose of permission
uses in custom code
 Location and contacts permission
 Key idea:
 Compiled Java code retain the text of many
identifiers, which offer a hint as to what the code is
doing (if not obfuscated)
void PhotoTag(){
…
loc = getLocation()
…
exif = readExif()
…
geotagging

How does the system work?
decompiled code:
…..
loc = getLocation(…)
…
keywords:
weather,
temperature,
wind,
….
Other features
classifier
App
Preprocessing
Feature
Extraction
Classification

Pre-processing
 Decompiling app
 Identify Permission-related code
 leverage the permission mapping provided by
PScout [CCS’12]
DEX
Smali
JAR Java
ApkTool
Dex2Jar
JD-JUI
Identify permission-
related code using
Smali code
Use Java code to
extract features

Feature Extraction(1)
Type Features Method
App-specific
features
Call frequency of each
permission-related API
calls, Intent, Content
Provider URIs
Static
Analysis
Text-based
features
Key words extracted
from identifiers
Text Mining

 App specific features
 Highly related to app behaviors
• e.g., API “sendTextMessage()” is often used for the “Call
and SMS” purpose
• e.g., APIs in the package “com.android.email.activity”
are often used for “email” purpose
 We use a list of 680 documented permission
related APIs, 97 Intent and 78 Content
Provider URI strings.

 Text-based features
 extract keywords from various identifiers as features
 Package/class/method/field names are preserved
when compiling
(1) Explicit patterns:
e.g.,
getHotel -> get, hotel
(2) Directory-based:
e.g.,
nearbyrestaurant ->
nearby, restaurant
Splitting Identifiers
Filter stop words:
e.g., get, set,
public, string
e.g.,
hotels -> hotel
restaurant -> restaur
Filtering Stemming TF-IDF
weather: 0.81236
wind: 0.76523
…

Classification
 Feature normalization
 Normalize features by scaling them to [0, 1]
 Supervised learning
 Manually label instances for each purpose
 A taxonomy of purposes
 10 purposes for the location permission and 10
purposes for the contacts permission

A taxonomy of purposes
Location permission
Search nearby places
Location-based
customization
Transportation information
Recording
Map and navigation
Geosocial networking
Geotagging
Location spoofing
Alert and remind
Location-based game
Contacts permission
Backup and Synchronization
Contact management
Blacklist
Call and SMS
Contact-based customization
Email
Find friends
Record
Fake calls and SMS
Remind

Evaluation
 Dataset
 7,923 apps from Google Play
• IDF is calculated based on a corpus of these apps
 Labeling purposes
 10 Cross-validation
1020 instances
560 instances
for contacts
460 instances
for location

Experiment Result
 Result of inferring purposes
Classifier Accuracy Precision Recall
Location
SVM 81.74% 85.51% 83.20%
Maximum Entropy 85.00% 87.07% 85.88%
C4.5 79.57% 83.26% 77%
Contacts
SVM 93.94% 94.38% 92.94%
Maximum Entropy 94.64% 94.42% 93.96%
C4.5 92.86% 91.36% 89.59%

Qualitative Analysis
 Determining factor:
 distinctive features, and the number of features
 Findings:
 Categories with high precision and recall tend to
have distinctive features.
 Most misclassified instances have fewer features

Feature comparison
 Text-based features achieve good accuracy alone
 App-specific features offer marginal
improvements
Permission Classifier Accuracy
(text features)
Accuracy
(all features)
Diff
Location Maximum
Entropy
81.97% 85.00% 3.03%
Contacts Maximum
Entropy
93.57% 94.64% 1.07%

Summary
 Develop a method to infer the purpose of
permission use in custom code
 Based on text-mining and machine-learning
 Evaluate our approach for two frequently used
permissions and could achieve high accuracy

Future work
 Infer the purpose and monitor app behaviors
at runtime
 Enforce access control based on purposes
 Automated infer the purpose of third-party
libraries
 Previous work only use manually labeled white-list

Take-away Message
 Knowing the purpose of
permission use is important
 An app could use a permission
for different purposes
 People have different level of concern for sensitive
data use for different purposes
 Text-mining on the decompiled code is useful to
infer the purpose of permission use
Thanks! Questions?
howiepku@pku.edu.cn
sei.pku.edu.cn/~wanghy11

Using Text Mining to Infer the Purpose of Permission Use in Mobile Apps

More Related Content

Viewers also liked

Similar to Using Text Mining to Infer the Purpose of Permission Use in Mobile Apps

Recently uploaded

Using Text Mining to Infer the Purpose of Permission Use in Mobile Apps