Memory Analysis of the Dalvik (Android) Virtual Machine
7 likes2,964 views
The document summarizes research on analyzing the memory of the Dalvik virtual machine used in Android. It describes acquiring memory from Android devices, locating key data structures in memory like loaded classes and their fields, and analyzing specific Android applications to recover data like call histories, text messages, and location information. The goal is to develop forensics capabilities for investigating Android devices through memory analysis.
![What is Dalvik?
• Dalvik is the software VM for all Android
applications
• Nearly identical to the Java Virtual Machine
(JVM) [1]
• Open source, written in C / Java
4](https://image.slidesharecdn.com/android-memory-analysis-111229114843-phpapp01/85/Memory-Analysis-of-the-Dalvik-Android-Virtual-Machine-4-320.jpg)
![Memory Forensics Introduction
• Memory forensics is vital to orderly recovery of
runtime information
• Unstructured methods (strings, grep, etc) are
brittle and only recover superficial info
• Structured methods allow for recovery of data
structures, variables, and code from memory
• Previous work at operating system level led to
recovery of processes, open files, network
connections, etc [4,5]
6](https://image.slidesharecdn.com/android-memory-analysis-111229114843-phpapp01/85/Memory-Analysis-of-the-Dalvik-Android-Virtual-Machine-6-320.jpg)
![Memory Analysis Process
• First, need to acquire memory
– Acquisition depends on environment [6]
• Next, requires locating information in memory
and interpreting it correctly
– Also requires re-implementing functionality offline
• Then it needs to be displayed in a useful way
to the investigator
7](https://image.slidesharecdn.com/android-memory-analysis-111229114843-phpapp01/85/Memory-Analysis-of-the-Dalvik-Android-Virtual-Machine-7-320.jpg)
![Acquiring Memory – Approach 2
• Memory can be acquired on a per-process
basis
• Ensures that all pages of the process will be
acquired
• Easiest to perform with memfetch[8]
– After a few small changes, was statically compiled
for ARM
• No unmapped pages will be recovered though
– Heap and GC don’t munmap immediately
11](https://image.slidesharecdn.com/android-memory-analysis-111229114843-phpapp01/85/Memory-Analysis-of-the-Dalvik-Android-Virtual-Machine-11-320.jpg)
![Gathering libdvm’s Structures
1) Grab the shared library from the phone (adb)
2) Use Volatility’s dwarfparse.py:
• Builds a profile of C structures along with
members, types, and byte offsets
• Records offsets of global variables
3) Example structure definition
'ClassObject': [ 0xa0, { Class name and size
'obj': [0x0, ['Object']], member name, offset,
and type
16](https://image.slidesharecdn.com/android-memory-analysis-111229114843-phpapp01/85/Memory-Analysis-of-the-Dalvik-Android-Virtual-Machine-16-320.jpg)
![Methods in Memory vs on Disk
• Dalvik makes a number of runtime
optimizations [1]
• Example: When class members are accessed
(iget, iput) the field table index is replaced
with the direct byte offset
• Would likely need to undo some of the
optimizations to get complete baksmali
output
25](https://image.slidesharecdn.com/android-memory-analysis-111229114843-phpapp01/85/Memory-Analysis-of-the-Dalvik-Android-Virtual-Machine-25-320.jpg)
![Finding Data Structures
• Can save substantial time by using adb’s
logcat (next slide)
– Shows the classes and often methods involved in
handling UI events
• Otherwise, need to examine source code
– Some applications are open source
– Others can be “decompiled” with baksmali [9]
28](https://image.slidesharecdn.com/android-memory-analysis-111229114843-phpapp01/85/Memory-Analysis-of-the-Dalvik-Android-Virtual-Machine-28-320.jpg)
![Implementation
• Recovery code written as Volatility [7] plugins
– Most popular memory analysis framework
– Has support for all Windows versions since XP and
2.6 Intel Linux
– Now also supports ARM Linux/Android
• Makes rapid development of memory analysis
capabilities simple
• Also can be used for analyzing other binary
formats
41](https://image.slidesharecdn.com/android-memory-analysis-111229114843-phpapp01/85/Memory-Analysis-of-the-Dalvik-Android-Virtual-Machine-41-320.jpg)
![Future Avenues of Research
• Numerous applications with potentially
interesting information
– Too much to manually dig through
– Need automation
– Baksmali/Volatility/logcat integration?
• Automated determination of interesting
evidence across the whole system
– Combing work done in [2] and [3]
44](https://image.slidesharecdn.com/android-memory-analysis-111229114843-phpapp01/85/Memory-Analysis-of-the-Dalvik-Android-Virtual-Machine-44-320.jpg)
![References - 1
[1] http://bit.ly/dalvikvsjava
[2] Brendan Dolan-Gavitt, et al, “Virtuoso: Narrowing
the Semantic Gap in Virtual Machine
Introspection”, IEEE Security and Privacy, 2011
[3] TaintDroid, http://www.appanalysis.org/
[4] http://bit.ly/windowsmemory
[5] http://bit.ly/linuxmem
[6] http://bit.ly/memimaging
[7] http://code.google.com/p/volatility/
[8] http://lcamtuf.coredump.cx/soft/memfetch.tgz
46](https://image.slidesharecdn.com/android-memory-analysis-111229114843-phpapp01/85/Memory-Analysis-of-the-Dalvik-Android-Virtual-Machine-46-320.jpg)
![References - 2
[9] baksmali - http://code.google.com/p/smali/
47](https://image.slidesharecdn.com/android-memory-analysis-111229114843-phpapp01/85/Memory-Analysis-of-the-Dalvik-Android-Virtual-Machine-47-320.jpg)
Memory Analysis of the Dalvik (Android) Virtual Machine
- 1. Memory Analysis of the Dalvik (Android) Virtual Machine Andrew Case Digital Forensics Solutions
- 2. Who Am I? • Security Analyst at Digital Forensics Solutions – Also perform wide ranging forensics investigations • Volatility Developer • Former Blackhat and DFRWS speaker 2
- 3. Agenda • What is Dalvik / why do we care? • Brief overview of memory forensics • Extracting allocated and historical data from Dalvik instances • Target specific Android applications 3
- 4. What is Dalvik? • Dalvik is the software VM for all Android applications • Nearly identical to the Java Virtual Machine (JVM) [1] • Open source, written in C / Java 4
- 5. Why do we care? • Android-based phones leading in US mobile market – Which makes for many phones to investigate • Memory forensics capabilities against Android applications have numerous uses/implications • Entire forensics community (LEO, .gov, private firms) already urging development of such capabilities 5
- 6. Memory Forensics Introduction • Memory forensics is vital to orderly recovery of runtime information • Unstructured methods (strings, grep, etc) are brittle and only recover superficial info • Structured methods allow for recovery of data structures, variables, and code from memory • Previous work at operating system level led to recovery of processes, open files, network connections, etc [4,5] 6
- 7. Memory Analysis Process • First, need to acquire memory – Acquisition depends on environment [6] • Next, requires locating information in memory and interpreting it correctly – Also requires re-implementing functionality offline • Then it needs to be displayed in a useful way to the investigator 7
- 9. Acquiring Memory – Approach 1 • The normal method is to acquire a complete capture of physical RAM • Works well when analyzing kernel data structures as their pages are not swapped out • Allows for recovery of allocated and historical processes, open files, network connections, and so on 9
- 10. Approach 1 on Android • Without /dev/mem support, need a LKM to read memory • No current module works for Android (ARM) • We developed our own (mostly by @jtsylve) • Benefits of full capture: – Can target any process (including its mappings) – Can recover information from unmapped pages in processes 10
- 11. Acquiring Memory – Approach 2 • Memory can be acquired on a per-process basis • Ensures that all pages of the process will be acquired • Easiest to perform with memfetch[8] – After a few small changes, was statically compiled for ARM • No unmapped pages will be recovered though – Heap and GC don’t munmap immediately 11
- 12. Analyzing C vs Java • Most previous forensics research has had the “luxury” of analyzing C – Nearly 1:1 mapping of code/data to in-memory layout • Declaration of a C “string” – char buffer*+ = “Hello World”; • Memory Layout (xxd) – 4865 6c6c 6f20 576f 726c 6400 Hello World. 12
- 13. A Dalvik String in Memory • First, need the address of the “StringObject” • Next, need the offsets of the “java/lang/String” value and byte offset members • StringObject + value offset leads us to an “ArrayObject” • ArrayObject + byte offset leads to an UTF-16 array of characters • … finally we have the string (in Unicode) 13
- 14. Now for the memory analysis… • The real goal of the research was to be able to locate arbitrary class instances and fields in memory • Other goals included replicating commonly used features of the Android debugging framework 14
- 15. Locating Data Structures • The base of Dalvik loads as a shared library (libdvm.so) • Contains global variables that we use to locate classes and other information • Also contains the C structures needed to parse and gather evidence we need 15
- 16. Gathering libdvm’s Structures 1) Grab the shared library from the phone (adb) 2) Use Volatility’s dwarfparse.py: • Builds a profile of C structures along with members, types, and byte offsets • Records offsets of global variables 3) Example structure definition 'ClassObject': [ 0xa0, { Class name and size 'obj': [0x0, ['Object']], member name, offset, and type 16
- 17. Volatility Plugin Sample • Accessing structures is as simple as knowing the type and offset intval = obj.Object(“int”, offset=intOffset, ..) • Volatility code to access ‘descriptor’ of an ‘Object’: o = obj.Object("Object", offset=objectAddress, ..) c = obj.Object("ClassObject", offset=o.clazz, …) desc = linux_common.get_string(c.descriptor) 17
- 18. gDvm • gDvm is a global structure of type DvmGlobals • Holds info about a specific Dalvik instance • Used to locate a number of structures needed for analysis 18
- 19. Locating Loaded Classes • gDvm.loadedClasses is a hash table of ClassObjects for each loaded class • Hash table is stored as an array • Analysis code walks the backing array and handles active entries – Inactive entries are NULL or have a pointer value of 0xcbcacccd 19
- 20. Information Per Class • Type and (often) name of the source code file • Information on backing DexFile – DexFile stores everything Dalvik cares about for a binary • Data Fields – Static – Instance • Methods – Name and Type – Location of Instructions 20
- 21. Static Fields • Stored once per class (not instance) • Pre-initialized if known • Stored in an array with element type StaticField • Leads directly to the value of the specific field 21
- 22. Instance Fields • Per instance of a Class • Fields are stored in an array of element type InstField • Offset of each field stored in byteOffset member – Relative offset from ClassObject structure 22
- 23. Listing Instance Members Source file: ComposeMessageActivity.java Class: Lcom/android/mms/ui/ComposeMessageActivity; Instance Fields: name: m_receiver signature: Landroid/content/BroadcastReceiver; name: m_filter signature: Landroid/content/IntentFilter; name: mAppContext signature: Landroid/content/Context; name: mAvailableDirPath signature: Ljava/lang/String; 23
- 24. Analyzing Methods • We can enumerate all methods (direct, virtual) and retrieve names and instructions • Not really applicable to this talk • Can be extremely useful for malware analysis though – If .apk is no longer on disk or if code was changed at runtime 24
- 25. Methods in Memory vs on Disk • Dalvik makes a number of runtime optimizations [1] • Example: When class members are accessed (iget, iput) the field table index is replaced with the direct byte offset • Would likely need to undo some of the optimizations to get complete baksmali output 25
- 27. Recovery Approach • Best approach seems to be locating data structures of UI screens – UI screens represented by uniform (single type) lists of displayed information – Data for many views are pre- loaded 27
- 28. Finding Data Structures • Can save substantial time by using adb’s logcat (next slide) – Shows the classes and often methods involved in handling UI events • Otherwise, need to examine source code – Some applications are open source – Others can be “decompiled” with baksmali [9] 28
- 29. logcat example The following is a snippet of output when clicking on the text message view: D/ConversationList(12520): onResume Start D/ComposeMessageActivity(12520): onConatctInfoChange D/RecipientList(12520): mFilterHandler not null D/RecipientList(12520): get recipient: 0 D/RecipientList(12520): r.name: John Smith D/RecipientList(12520): r.filter() return result D/RecipientList(12520): indexOf(r)0 D/RecipientList(12520): prepare set, index/name: 0/John Smith 29
- 30. Phone Call History • Call history view controlled through a DialerContactCard$OnCard ClickListener • Each contact stored as a DialerContactCard • Contains the name, number, convo length, and photo of contact 30
- 31. Per Contact Call History • Can (sometimes) retrieve call history per- contact • Requires the user to actually view a contact’s history before being populated 31
- 32. Text Messages • Recovery through ComposeMessageActivity & TextMessageView • Complete conversations can be recovered • Not pre-populated 32
- 33. Voicemail • Audio file is open()’ed • Not mapped contiguously into the process address space • No method to recover deleted voicemails.. 33
- 34. Browser (Opera Mini) • Opera Mini is the most used mobile browser • Can recover some session information – The history file is always mapped in memory (including information from current session) • HTTP requests and page information is (possibly) recoverable – Can recover <title> information – Stored in Opera Binary Markup Language – Not publicly documented? 34
- 35. Recovering Wireless Information • Screenshot on the right shows results of a scan for wireless networks • Recovery of this view provides the SSID, MAC address, and enc type for routers found • Recovery of “Connected” routers show which were associated with 35
- 36. Other Wireless Information • Potentially interesting information: – Wireless keys – Connection stats • These are not controlled by Dalvik – Keys only initially entered through Dalvik, but then saved • Stored by the usual Linux applications – wpa_supplicant, dhcpd, in-kernel stats 36
- 37. Location Recovery • Associating location & time not always important – But makes for better slides *hint* • Interesting for a number of reasons – Forensics & Privacy concerns – Not part of a “standard” forensics investigation 37
- 38. Google Maps • Did not do source code analysis – Most phones won’t be using Google Maps while being seized – Wanted to find ways to get historical data cleanly • Found two promising searches – mTime=TIME,mLatitude=LAT,mLongitude=LON – point: LAT,LON … lastFix: TIME • TIME is the last location, extra work needed to verify 38
- 39. “Popular” Weather Application • The weather application uses your location to give you relevant information • http://vendor.site.com/widget/search.asp? lat=LAT&lon=LON&nocache=TIME 39
- 40. More GPS Fun • All of the following applications do not clear GPS data from memory, and all send their lat/lon using GET with HTTP – Urban Spoon – Weather Channel – WeatherBug – Yelp – Groupon – Movies 40
- 41. Implementation • Recovery code written as Volatility [7] plugins – Most popular memory analysis framework – Has support for all Windows versions since XP and 2.6 Intel Linux – Now also supports ARM Linux/Android • Makes rapid development of memory analysis capabilities simple • Also can be used for analyzing other binary formats 41
- 42. Testing • Tested against a HTC EVO 4G – No phone-specific features used in analysis – Only a few HTC-specific packages were analyzed • Visually tested against other Dalvik versions – No drastic changes in core Dalvik functionality 42
- 43. Research Applications • Memory forensics (obviously) • Testing of privacy assurances • Malware analysis – Can enumerate and recover methods and their instructions 43
- 44. Future Avenues of Research • Numerous applications with potentially interesting information – Too much to manually dig through – Need automation – Baksmali/Volatility/logcat integration? • Automated determination of interesting evidence across the whole system – Combing work done in [2] and [3] 44
- 46. References - 1 [1] http://bit.ly/dalvikvsjava [2] Brendan Dolan-Gavitt, et al, “Virtuoso: Narrowing the Semantic Gap in Virtual Machine Introspection”, IEEE Security and Privacy, 2011 [3] TaintDroid, http://www.appanalysis.org/ [4] http://bit.ly/windowsmemory [5] http://bit.ly/linuxmem [6] http://bit.ly/memimaging [7] http://code.google.com/p/volatility/ [8] http://lcamtuf.coredump.cx/soft/memfetch.tgz 46
- 47. References - 2 [9] baksmali - http://code.google.com/p/smali/ 47