200 likes | 318 Views
Behavioral Detection of Malware on Mobile Handsets MobiSys 2008, Abhijit Bose et al. 2008.12.11. 이 승 민. Contents. 1. Introduction. 2. System Overview. 3. Malicious Behavior Signatures. 4. Run-time Construction. 5. Evaluation. 6. Conclusion. 1. Introduction. Behavior ?.
E N D
Behavioral Detection of Malware on Mobile Handsets MobiSys 2008, Abhijit Bose et al. 2008.12.11. 이 승 민
Contents 1 Introduction 2 System Overview 3 Malicious Behavior Signatures 4 Run-time Construction 5 Evaluation 6 Conclusion
1. Introduction • Behavior ?
1. Introduction • Malware on mobile handsets • The first mobile worm Cabir appeared in June 2004 • By the end of 2006, the known number of mobile malware families and variants increased by 59% and 75% from year 2005 • Differences in mobile • Limited resources such as CPU, memory and battery • Difficulty of constructing network signature • Spreading via non-traditional vectors (SMS, Bluetooth) • Difference in OS (file permission, modification)
1. Introduction • Related work • Network based anomaly detection • Host based anomaly detection • Using consecutive system calls from normal app. • Rule learning, finite-state automata, Hidden Markov Model • But, it could be evaded by simple obfuscation • This paper • Monitoring a program run-time behavior at a higher level • Run-time analysis • Using both normal and malware behaviors
2. System Overview • System • Monitor agent collects the application behavior in the form of system events/API calls • Aggregated behavior signatures are reported to the detection agent
3. Malicious Behavior Signatures • Temporal patterns • A logical ordering of the steps over time often clearly reveals the malicious intent • Example • Bluetooth OBEX system call (CObexClient::Put()) Harmless • Received file is of type .SIS & that file is later executed & the installer process seeks to overwrite files in the system directory Mabir, Commwarrior • Behavior signatures are best specified using temporal logic instead of classical propositional logic • TLCK (temporal logic of causal knowledge) language
3. Malicious Behavior Signatures • Temporal logic • Specify malicious behavior in terms of system events, by temporal and logical operators • : true at time t • : true at some instant before t • : true at all instants before t • : true at some instant in the interval [t-k, t]
3. Malicious Behavior Signatures • Example: Commwarrior Worm • Symbian OS • Atomic variables are combined into seven higher-level signatures
4. Run-Time Signature Construction • Monitoring API Calls using Proxy DLL • Proxy DLL intercepts and records details about the API call events from the application (with Symbian OS emulator)
4. Run-Time Signature Construction • Stage I: Generation of dependency graph • Dependency graph is constructed from logged API calls
4. Run-Time Signature Construction • Stage II: Graph pruning and aggregation • Dependency graph grows over time • Pruning • The process did not have inter-process dependency relationships with any other process • Its graph does not partially match with any malicious behavioral signatures • It did not create or modify any file or directory • It is a helper process that takes input from a process and returns data to the main process • Aggregation • Each API call is aggregated to reduce the size of the overall storage • Construction of a behavior signature (TLCK)
5. Evaluation • SVM classification • Which of the separators is optimal ?
5. Evaluation • Margin of the separator is the width of separation between classes • Maximizing the margin is good according to intuition • Examples closest to the hyperplane are support vectors
5. Evaluation • Methodology • Monitoring agent is implemented in the Symbian OS Emulator • OS dependent • 8 applications • 5 worms: Cabir, Mabir, Lasco, Commwarrior, generic worm • 3 legitimate: OBEX file transfer, MMS client, MakeSIS • Detection agent uses SVM classifier • OS independent
5. Evaluation • Accuracy of SVM • Detection for known worms • SVM almost never falsely classifies a legitimate application signature to malicious
5. Evaluation • Detection for unknown worms • When the training set contains 3 malware, detection is relatively high
6. Conclusion • Contribution • First attempt to construct a behavioral detection model for mobile environments • Define malicious behaviors with TLCK (temporal logic) • Discussion • What is the difference compared to wired network? • How about using HMM (Hidden Markov Model) in behavior detection? • Suitable for future research topic?