Thesis Proposal Slides

Presentation slides from my PhD proposal defense

Transcript

1. A Deep and Longitudinal Approach to Mining Mobile Applications Khalid Alharbi Sikuli Lab University 
 of Colorado
 Boulder Thesis Proposal November 30th, 2015
2. Outline •Introduction •Background •Approach •Mining Design Changes •Deep Search Engine •Deep Search Queries •Evaluation
4. Mobile apps have become a valuable data source to mine and extract insights from “ ”
5. Mining Store Listing Details Data - Mining reviews to perform sentiment analysis and summarize users’ feedback. - Mining specific listing information such as description, maturity ratings, what’s new, category, developer, etc.
6. Mining User Interface Data - Mining User Interface Layout files to find design examples and compute UI complexity. - Mining UI text data to identify possible stealthy behavior.
7. Mining Code Data - Applying static analysis to study permission and API usages.
 - Detecting patterns similar to malicious applications. - Detecting privacy violations.
8. Previous studies are shallow and mostly focus on a single release of the app. “ ”
10. UI Design Analysis public class MainActivity extends AppCompatActivity { } <android.support.v7.widget.Toolbar android:id="@+id/my_toolbar" android:theme=“@style/ThemeOverlay.AppCompat.ActionBar" /> <application android:theme=“@style/Theme.AppCompat.Light.NoActionBar"/> AndroidManifest.xml MainActivity.java main_activity.xml Code UI Manifest
11. Sentiment Analysis Code UI
12. Privacy Analysis
13. Security Analysis
14. Deep Approach Deep and structural indexing of apps across multiple levels (listing, UI, code, libraries, backend providers) “ ”
15. Longitudinal Approach Observe the dynamics of apps over long time across multiple levels “ ”
16. Thesis Statment The goal of this work is to enable and advance mining mobile apps at deep and longitudinal levels to gain insights into various aspects of mobile apps at large-scale “ ”
17. Research Questions •RQ1: What are the benefits of the deep and longitudinal approach to mining mobile apps? •RQ2: How can we enable deep searching and mining of mobile apps over time?
18. RQ1.1: What could a deep and longitudinal approach achieve in UI design mining ? Collect, Decompile, Extract, Stats, and Diff: Mining Design Pattern Changes in Android Apps Khalid Alharbi and Tom Yeh, MobileHCI 2015, Honorable Mention Award
23. UI Design Patterns are always evolving How can we quantify the changes?
24. Data-driven Approach •Collect apps and their subsequent versions •Decompile apps to their nearly original form •Extract a comprehensive set of features •Stats compute statistics on extracted features •Diff compute the difference between app versions
25. Collect •18 months •24,436 unique apps •At least 2 versions per app •In total 56,349 APKs
27. Decompile
28. Features Extraction Listing Details Title, description, reviews, category, price, date published, ratings count, rating, install size, downloads count, permissions, what’s new. Appearance Layout directories, layout files, view group containers, view elements, relationships, drawable resources, UI text resources. Behavior app framework invocations, manifest (AndroidManifest.xml), third-party libraries.
29. Stats •Compute descriptive statistics •What is the percentage of apps that used a given design pattern? •What is the percentage of apps that switched into an alternative design pattern? •What is the percentage of apps that maintained the use of a given design pattern?
30. Diff •Detect changes that occur to the extracted features •See how changes are reflected
31. Analysis System
32. Application •Navigation design patterns •Tab Layout Fragments •Horizontal Paging Action Bar •Up Navigation Navigation Drawers •Custom UI Components •Homescreen widgets
33. Tabs with TabHost •Introduced in the initial release of the Android UI framework. •Deprecated in API level 11 in favor of new navigation patterns, such as the Fragment and the Action Bar patterns.
34. Tabs with TabHost •3,809 (15.6%) apps used TabHost •666 (17.5%) of them used it for the first time •Migration rate is very slow •413 (10.8%) apps stopped using it and switched to other types of navigations •New apps are using it
35. Before After
36. Navigation Drawer
37. Navigation Drawer •1,183 (4.8%) apps used Navigation Drawer •771 (65.2%) of them used it for the first time •Adoption rate is very slow •37 (3.1%) apps stopped using it and switched to other types of navigations
38. Before After
39. Findings: Some Highlights •Some apps would switch to a design pattern even after it was deprecated. •The adoption rate of newly introduced design patterns is relatively low. •Some apps would update their listing details to reflect changes in design patterns. •Some design patterns have slow adoption rate but higher continuity of use over multiple releases.
40. Limitations •Not deep enough •Too specific; requires specialized scripts •Not fully automated •No longitudinal study
41. LOC = 2,421 Does not scale well
42. RQ2: How can we enable deep searching and mining of mobile apps over time? RQ1: What are the benefits of the deep and longitudinal approach to mining mobile apps?
43. A Deep Search Engine 
 Sieveable: Deep Android Apps Search Khalid Alharbi, Jackson Chen, and Tom Yeh, VLDB 2016 (in submission)
44. Challenges • Dealing with heterogeneous data • Store Listing Details: document-oriented • UI: hierarchical tree structure • Code: text data • Creating structural indexes • Designing an effective and generalizable search engine.
45. Requirements • Deep • Generalizable • Extensible • Scalable
46. Data Collection • Web and marketplace crawlers • Collected over 400,000 apps • Multiple versions are collected • Decompile downloaded apps • Extract listing details, UI, manifest, and code features • Total dataset size is over10 TB
47. Features Indexing • Listing details • Single-field and text indexes • UI • Structural index that keeps track of all DOM elements’ relationships. • Manifest •Index DOM elements and their attributes • Code •Full-text indexes
48. Indexing UI data • Tag Index • Holds tag names and their attributes • • Structural Index •A suffix tree based index that stores the XML tree in a suffix array format RelativeLayout->LinearLayout-> ImageButton EditText(android: layout_width=“fill_parent”)
49. Query Language Specification • SQL-like declarative query syntax MATCH app WHERE <developer>Google Inc.</developer> <LinearLayout> <Button/> </LinearLayout> <uses-permission android:name= “android.permission.SEND_SMS"/> <code class=“android.hardware.Camera" method="takePicture"/> RETURN app
50. Query Language Specification • Query Parser MATCH app WHERE <developer>Google Inc.</developer> <LinearLayout> <Button/> </LinearLayout> <uses-permission android:name= “android.permission.SEND_SMS"/> <code class=“android.hardware.Camera" method="takePicture"/> RETURN app Listing UI Manifest Code
51. Query Language Specification • Query Executor Listing Details Code Manifest Index Manifest DOM UI Index UI DOM MATCH app WHERE <developer>Google Inc.</developer> <LinearLayout> <Button/> </LinearLayout> <uses-permission android:name= “android.permission.SEND_SMS"/> <code class=“android.hardware.Camera" method="takePicture"/> RETURN app 125K apps 249 apps 66 apps 21 apps 21 apps 15 apps 15 apps
52. System Architecture
53. Deep Search Queries •UI Design •Program Analysis •Security
54. UI Search Queries MATCH app WHERE <store-category>(*)</store-category> <PreferenceScreen> <ListPreference/> </PreferenceScreen> <code class="android.content.Context" method="getSharedPreferences"/> RETURN app, $1 Find the store category of apps with Settings/Preferences Screens
55. Program Analysis Queries MATCH app WHERE <code class="android.hardware.Camera" method="open"/> <uses-permission android:name="android.permission.CAMERA"/> RETURN app MATCH app WHERE <code class="android.app.Service" method="onStartCommand"/> <whats-new>bug fixes</whats-new> RETURN app • Overprivilege Analysis • Bug Fixes Analysis
56. Security Queries MATCH app WHERE <code method="addJavascriptInterface"/> <uses-sdk android:targetSdkVersion="12"/> RETURN app MATCH app WHERE <uses-permission android:name="android.permission.*" __min="11" /> <uses-permission android:name="*_SMS"/> RETURN app • Potential software vulnerability • Permissions
57. Evaluation •Evaluation Metrics •Evaluation of the Approach
58. Evaluation Metrics •Effectiveness •Precision and Recall •Benchmarks •Indexing Time •Query Execution Time
59. Approach Evaluation •[Deep] As the number of levels increases, the accuracy of the analysis increases significantly. •[Longitudinal] As the number of versions increases, the more insights we obtain about mobile apps.
60. Evaluation: Deep Approach •Comparison against single view approaches. •Find apps with material design lists and cards. •Randomly select 1000 apps •Inspect screenshots and description in listing •Inspect their UI layout files •Inspect their bytecode files •As we go deeper in the analysis, both precision and recall increase significantly.
61. Evaluation: Longitudinal Approach •Evolution and adoption rate experiments •Example: Find the time period that has seen a spike in adopting material design •Collect ground truth: use web search engines to collect 100 number of apps that have switched to material design. •Use Sieveable to find the release date of the first adopted version and compare the result with the ground truth.
62. Contributions • A novel approach that enables mining mobile apps at multiple levels over time manifested in a search engine called Sieveable • Demonstrated the utility of the approach by conducting diverse types of analyses •The proposed approach has been implemented, deployed, and people can actually use it
63. TimeLine Task Date Evaluation December 2015 - January 2016 Writing and completing dissertation February 2016 Defending the dissertation March 2016
64. A Deep and Longitudinal Approach to Mining Mobile Applications Khalid Alharbi Tom Yeh, Jackson Chen, Sanghee Kim, and everyone at the Sikuli lab. King Abdulaziz University. everyone else who is here today. & Thank you & http://sieveable.io These slides are available at:
 https://speakerdeck.com/kalharbi