TKPERM: Cross-platform Permission Knowledge Transfer to Detect Overprivileged Third-party Applications Faysal Hossain Shezan, Kaiming Cheng, Zhen Zhang, Yinzhi Cao, Yuan Tian
Permission-based Access Control Android Chrome IFTTT 2
Permission Correlation with Description Android App 3 UBER- https://play.google.com/store/apps/details?id=com.ubercab&hl=en_US
Permission Correlation with Description Location Permission Android App Requested Permission 4 UBER- https://play.google.com/store/apps/details?id=com.ubercab&hl=en_US
Permission Correlation with Description The app uses your location so your Location Permission driver knows where to pick you up. Android App Requested Permission Uber Description 5 UBER- https://play.google.com/store/apps/details?id=com.ubercab&hl=en_US
Permission Correlation with Description The app uses your location so your Location Permission driver knows where to pick you up. Consistent Android App Requested Permission Uber Description 6 UBER- https://play.google.com/store/apps/details?id=com.ubercab&hl=en_US
What is Overprivileged? GamingHub (Chrome Extension) 7 GamingHub- https://chrome.google.com/webstore/detail/gaminghub/eafoaklfmpnpdecnhhaailihkdbhkgin
What is Overprivileged? GamingHub Location Permission (Chrome Extension) Requested Permission 8 GamingHub- https://chrome.google.com/webstore/detail/gaminghub/eafoaklfmpnpdecnhhaailihkdbhkgin
What is Overprivileged? Primary Features: 1. Quick & Easy Access to popular web games 2. Minimalist & Elegant Design 3. Hand Picked High Quality GamingHub Wallpapers that change according Location Permission (Chrome Extension) to mood 4. New & Exciting ways for accessing Online Content 5. Let us know what you'd like, more to come soon! Requested Permission GamingHub Description 9 GamingHub- https://chrome.google.com/webstore/detail/gaminghub/eafoaklfmpnpdecnhhaailihkdbhkgin
What is Overprivileged? No Explanation for the GamingHub Usage of Location Location Permission (Chrome Extension) Permission No Match Requested Permission GamingHub Description 10 GamingHub- https://chrome.google.com/webstore/detail/gaminghub/eafoaklfmpnpdecnhhaailihkdbhkgin
Challenges 11 Taken from: https://iot-analytics.com/iot-platform-companies-landscape-2020/
Challenges Extensive data labeling and parameter tuning on new platforms Some platforms have limited data 12 Taken from: https://iot-analytics.com/iot-platform-companies-landscape-2020/
Key Insights Permission Knowledge Location Android App Chrome App 13
Solution- Transfer Learning 14
Goal General framework to detect unexpected permissions 15
Research Questions 1. What knowledge to transfer? (e.g., what original domain should we select, what permissions in Android should we use)? 2. How to minimize the amount of labeled data needed? 16
System Overview of TKPERM Source Platform Read Contacts Access Coarse Location Access Fine Location ………. Camera 17
System Overview of TKPERM Source Platform Read Contacts Access Coarse Location 1 Access Fine Location Domain ………. Selection Camera 18
System Overview of TKPERM Source Platform Read Contacts Source Model Training Access 2 Coarse Location 1 Access Fine Location Domain ………. Selection Camera 19
System Overview of TKPERM 3 Source Platform Read Contacts Source Model Source Model Training Access 2 Coarse Location 1 Access Fine Location Domain ………. Selection Camera 20
System Overview of TKPERM 3 Target Platforms Source Platform Chrome Read Contacts Source Model Geolocation Source Model Training Access Chrome Proxy 2 Coarse Location 1 Chrome Content Access Settings Fine Location Domain ………. ………. Selection SmartThings Camera Switch 21
System Overview of TKPERM 3 Target Platforms Source Platform + 4 5 Chrome Read Contacts Source Model Geolocation Source Model Training Access Chrome Proxy 2 Coarse Location Data Selection 1 Chrome Content Access Settings Fine Location Domain ………. ………. Selection SmartThings Camera Switch 22
System Overview of TKPERM 3 Target Platforms Source Platform + 4 5 Chrome Read Contacts Source Model Geolocation Source Model Training Access Chrome Proxy 2 Coarse Location Data Selection + 7 6 1 Chrome Content Access Settings Fine Location Domain ………. ………. Selection SmartThings Camera Target Model Switch Training 23
System Overview of TKPERM 3 Target Platforms Source Platform + 4 5 Chrome Read Contacts Source Model Geolocation Source Model Training Access Chrome Proxy 2 Coarse Location Data Selection + 7 6 1 Chrome Content Access Settings Fine Location 8 Domain ………. ………. Selection SmartThings Camera Target Model Switch Target Model Training 24
System Overview of TKPERM 3 Target Platforms Source Platform + 4 5 Chrome Read Contacts Source Model Geolocation Source Model Training Access Chrome Proxy 2 Coarse Location Data Selection + 7 6 1 Chrome Content Access Settings Fine Location 8 Domain ………. ………. Selection SmartThings Camera Target Model Switch Target Model Training 25
Domain Selection Research Question: What knowledge to transfer? Greedy Selection Approach Compute and aggregate source domain(s) performs 26
Domain Selection Research Question: What knowledge to transfer? Greedy Selection Approach Compute and Remove source aggregate source domain(s) which domain(s) work worst performs 27
Domain Selection Research Question: What knowledge to transfer? Greedy Selection Approach Compute and Find the best Remove source aggregate source combination of domain(s) which domain(s) the source work worst performs domain(s) 28
Domain Selection Greedy Selection Approach Compute and Find the best Remove source aggregate source combination of domain(s) which domain(s) the source work worst performs domain(s) Research Question: What knowledge to transfer? 29
Data Selection Research Question: How to minimize the amount of labeled data needed? Use source model to rank the unlabeled document 30
Data Selection Research Question: How to minimize the amount of labeled data needed? Use source Pick the top 20 model to rank documents from the unlabeled the target document domain 31
Data Selection Research Question: How to minimize the amount of labeled data needed? Use source Pick the top 20 Ask human model to rank documents from annotator to the unlabeled the target label data document domain 32
Data Selection Use source Pick the top 20 Ask human model to rank documents from annotator to the unlabeled the target label data document domain Research Question: How to minimize the amount of labeled data needed? 33
Dataset 36,193 Sentences 4,705 Sentences Android Chrome SmartThings IFTTT 292 666 Sentences Sentences 34 Available at: https://drive.google.com/drive/u/1/folders/1Yfnz-ZpBpL8lftYIdM6JtH-QKE88NcSX
Dataset 36,193 Sentences 4,705 Sentences Android Chrome SmartThings IFTTT 292 666 Sentences Sentences AUTOCOG 35 AutoCog: Measuring the Description-to-permission Fidelity in Android Applications, Qu et al. (CCS 2014)
Evaluation Question 1. What is the end-to-end performance of TKPERM? Question 2. What is the performance of each component in TKPERM? Question 3. What is the computation overhead of TKPERM? 36
Evaluation Question 1. What is the end-to-end performance of TKPERM? Question 2. What is the performance of each component Effectiveness in TKPERM? Question 3. What is the computation overhead of TKPERM? 37
Evaluation (Effectiveness) Source Domain Selection : H-divergence v/s Greedy Selection in IFTTT Platform Target Domain Source Selection Source Domain(s) F1 Evernote H-Divergence Read Calendar 75.86% Greedy Selection Coarse Location + Fine Location + Camera 83.13% BMW Lab H-Divergence Read Contact 92.30% Greedy Selection Send SMS + Record Audio 95.24% Facebook H-Divergence Read Calendar 76.09% Greedy Selection Camera 88.09% Google Calendar H-Divergence Read Calendar 91.30% Greedy Selection Read Calendar + Coarse Location 92.30% Google Contact H-Divergence Read Contacts 99.20% Greedy Selection Read Contacts 99.20% 38
Evaluation (Effectiveness) Source Domain Selection : H-divergence v/s Greedy Selection in IFTTT Platform Target Domain Source Selection Source Domain(s) F1 Evernote H-Divergence Read Calendar 75.86% Greedy Selection Coarse Location + Fine Location + Camera 83.13% BMW Lab H-Divergence Read Contact 92.30% Greedy Selection Send SMS + Record Audio 95.24% Facebook H-Divergence Read Calendar 76.09% Greedy Selection Camera 88.09% Google Calendar H-Divergence Read Calendar 91.30% Greedy Selection Read Calendar + Coarse Location 92.30% Google Contact H-Divergence Read Contacts 99.20% Greedy Selection Read Contacts 99.20% 39
Recommend
More recommend
