10 - Android formats¶
This tutorial introduces Android formats as well as the API to use them. We are talking about DEX, OAT, VDEX and ART.
Files used in this tutorial are available on the tutorials repository
By Romain Thomas - @rh0main
Let’s start with a quick reminder about compilation, installation and the execution of Android applications.
When dealing with application development, the main part of the code is usually written in Java. Developers can also write native code (
the Java Native Interface (JNI) interface.
In the APK building process, the Java code is eventually transformed in the Dalvik bytecode which is interpreted by the Android Java virtual machine. The Android JVM is different from the implementation by Oracle and, among the differences, it is based on registers whereas the one from Oracle is based on a stack.
To produce the Dalvik bytecode, Java sources are first compiled with
javac into the Java bytecode and then Android transforms this
bytecode into the Dalvik one by using the
dx compiler (or the new one:
D8). This bytecode is finally wrapped in a DEX file(s) such as
classes.dex. The DEX format is specific
to Android and the documentation is available here.
During the installation of the APK, the system applies optimizations on this DEX file in order to speed-up the execution. Indeed interpreting bytecode is not as efficient as executing native code and the Dalvik virtual machine is based on registers that are 32-bits width size whereas most of the recent CPU are 64-bits width.
To address this issue and prior to Android 4.4 (KitKat), the runtime used JIT compilation to transform Dalvik bytecode into assembly. The JIT ocurred during the execution and it was done each time the application was executed. Since Android 4.4 they moved to a new runtime which, among other features, performs the optimizations during the installation. Consequently the installation takes more time but transformations to native code are done once.
To optimize the Dalvik bytecode, the original DEX file (e.g.
classes.dex) is transformed into another file that will contain the native code. This new file usually has
.oat extension and is wrapped by the ELF format. Using ELF format makes sense for mainly two reasons:
- It’s the default format used by Linux and Android to package assembly code.
- It enables to use the same loader:
OAT files are in fact ELF and this is why, we choose to add this new format in LIEF. This ELF format is actually used as a wrapper over another format which is specific to Android: the OAT format.
Basically the ELF associated exports few symbols:
import lief oat = lief.parse("SomeOAT") for s in oat.dynamic_symbols: print(s)
oatdata OBJECT GLOBAL 1000 1262000 oatexec OBJECT GLOBAL 1263000 10d4060 oatlastword OBJECT GLOBAL 233705c 4 oatbss OBJECT GLOBAL 2338000 f5050 oatbsslastword OBJECT GLOBAL 242d04c 4
These symbols are a kind of pointers to specific part of the OAT format. For example,
oatdata will point to the begining of the underlying OAT format whereas
to the native code. For those who are interested in a deeper understand of OAT internal structures, See:
These different formats can be a bit confusing and to summarize:
DEX are transformed into
.odex files that are primarily ELF files wrapping a custom OAT format.
The structure of the OAT is poorly documented and its internal structures change for each version of Android without backward compatibility. It means that OAT files produced on Android 6.0.1 can only be used on this version.
In the Android framework the
dex2oat executable is responsible to convert and optimize the APK DEX files into OATs. This executable is located in the
/system/bin/ directory and we can have its output through logcat:
$ adb logcat -s "dex2oat:I" ... 05-04 10:16:37.218 1987 1987 I dex2oat : /system/bin/dex2oat --compiler-filter=speed --dex-file=/data/user/0/com.google.android.gms/snet/installed/snet.jar --oat-file=/data/user/0/com.google.android.gms/snet/dalvik-cache/snet.dex 05-04 10:16:37.688 1987 1998 W dex2oat : Compilation of void com.google.android.snet.Snet.enterSnetIdle(android.content.Context, android.os.Bundle) took 116.995ms 05-04 10:16:37.768 1987 1987 I dex2oat : ---------------------------------------------------- 05-04 10:16:37.768 1987 1987 I dex2oat : <SS>: S T A R T I N G . . . 05-04 10:16:37.768 1987 1987 E dex2oat : <SS>: oat location is not valid /data/user/0/com.google.android.gms/snet/dalvik-cache/snet.dex 05-04 10:16:37.768 1987 1987 I dex2oat : dex2oat took 552.045ms (threads: 8) arena alloc=3MB java alloc=1150KB native alloc=8MB free=3MB 05-04 12:25:50.878 10460 10460 I dex2oat : /system/bin/dex2oat --compiler-filter=speed ...
The output above is the transformation of SaftyNet DEX files, located in
/data/user/0/com.google.android.gms/snet/installed/snet.jar, into an OAT saved in
One can notice that the extension is
.dex so it should be a DEX file and not an OAT. Actually if we check the type:
$ file snet.dex snet.dex: ELF 64-bit LSB shared object, ARM aarch64, version 1 (GNU/Linux), dynamically linked, stripped
We can see an ELF.
Do not trust extensions: .dex can be DEX or OAT, .odex are OAT, .oat are OAT, …
The process of converting Java sources into the OAT can be simplified with the following diagram:
The Missing DEX¶
If we analyze applications from the Google PlayStore, we usually have the
classes.dex file(s) in APK. As this file contains the Dalvik bytecode, most of the tools rely
on this file to perform the analysis (decompilation, static analysis, …)
However when analyzing constructor firmwares (or ROM) these DEX files could miss. For example, if we are interested in
com.android.settings from Samsung,
the application is associated with the
/system/priv-app/SecSettings2 directory which has the following structure:
$ tree system/priv-app/SecSettings2 ├── oat │ └── arm64 │ └── SecSettings2.odex └── SecSettings2.apk 2 directories, 2 files
By looking at the files in
SecSettings2.apk, we can’t find
$ unzip -l ./SecSettings2.apk|grep -c "classes.dex" 0
Next to the
SecSettings2.apk, we find
SecSettings2.odex which is the OAT file resulting of the optimization of the missing DEX file.
As ROM developers control the Android version and the target architecture, they just have to provide the OAT file.
They can also use this “feature” to avoid analysis and reverse engineering of the application. As the Dalvik bytecode is located in the DEX file, without this file the analysis is quite limited.
Thankfully there is a copy of the original DEX within the OAT! Actually it’s not an exact copy as
dex2oat replaces some Dalvik instructions (like
invoke-virtual) with optimized ones  but starting from Android N, we can also recover the original instructions.
Prior Android Oreo (8.0.0) DEX files were embedded in the OAT itself and after Oreo, the transformation performed by
dex2oat generates two files:
- classes.odex: OAT containing native code
- classes.vdex: VDEX file containing copy of original DEX files
The DEX files originally located in the OAT has been exported in a new file with a new format: the VDEX format. This new format is completely different from OAT, especially it’s not an ELF.
In the same way as OAT format, VDEX internal structures change for each version of Android without backward compatibility.
It also exists tools    to extract DEX from OAT/VDEX files but the extraction  is either limited to OAT  or to VDEX . With LIEF we aim to provide a single framework to deal with these formats.
OAT and VDEX¶
As explained in the previous part, internal structures of the formats change for each version of Android. LIEF provides an abstraction of these modifications and the user can deal with OAT or VDEX without carrying of the underlying version of the OAT.
It currently supports OAT files from Android 6.0 Marshmallow (OAT v64) to Android 8.0.1 Oreo (OAT v131).
The OAT version is available with the
>>> import lief >>> lief.OAT.version("classes.odex") # From Android 6 64 >>> lief.OAT.version("classes.odex") # From Android 7 88
One can also access to the associated Android version by using
>>> lief.OAT.android_version(64) ANDROID_VERSIONS.VERSION_601 >>> lief.OAT.android_version(124) ANDROID_VERSIONS.VERSION_800 >>> lief.Android.code_name(lief.Android.ANDROID_VERSIONS.VERSION_800) 'Oreo' >>> lief.Android.version_string(lief.Android.ANDROID_VERSIONS.VERSION_800) "8.0.0"
>>> import lief >>> oat = lief.parse("classes.odex") >>> type(oat) _pylief.OAT.Binary >>> isinstance(oat, lief.ELF.Binary) True
Thus the same ELF API is available: adding sections, modifying dynamic entries, etc and the
lief.OAT.Binary object adds the following methods:
OAT binary representation
- get_class(self: _pylief.OAT.Binary, class_name: str) -> LIEF::OAT::Class
Classfrom its name
- get_class(self: _pylief.OAT.Binary, class_index: int) -> LIEF::OAT::Class
Classfrom its index
Check if the class if the given name is present in the current OAT binary
If the given OAT targets Android Marshmallow or Nougat (6 or 7) then DEX files can be retrieved with the
>>> len(oat.dex_files) # > 1 if multi-dex 1 >>> dex = oat.dex_files >>> dex.save("/tmp/classes.dex")
From the code above, the
lief.DEX.File has been extracted to
/tmp/classes.dex (with de-optimization).
If the given OAT targets Android Oreo or above, then extraction is done by using the VDEX file. The
lief.OAT.parse() function accepts an OAT file or
an OAT and a VDEX file. By providing the VDEX file in the second parameter, the
lief.OAT.Binary object will have the same functionalities as the one for OAT pre-Oreo.
If the VDEX file is not provided then
lief.OAT.Binary will have limited information:
# Without VDEX file >>> oat_oreo = lief.parse("KeyChain.odex") >>> len(oat_oreo.dex_files) 0 >>> len(oat_oreo.classes) 0 >>> len(oat_oreo.oat_dex_files) 1 >>> oat_dex_file = oat_oreo.oat_dex_files >>> print(oat_dex_file) /system/app/KeyChain/KeyChain.apk - (Checksum: 0x206c8ab1)
# With VDEX file >>> oat_oreo = lief.OAT.parse("KeyChain.odex", "KeyChain.vdex") >>> len(oat_oreo.dex_files) 1 >>> len(oat_oreo.classes) 17 >>> oat_oreo.dex_files.save("/tmp/classes.dex")
We can also use the LIEF’s VDEX module directly:
>>> vdex = lief.VDEX.parse("KeyChain.vdex")
As the VDEX format is completely different from OAT, ELF, PE and Mach-O the VDEX parser creates a
lief.VDEX.File object and not a
We can also extract DEX files with the
>>> len(vdex.dex_files) 1 >>> vdex.dex_files.save("/tmp/KeyChain.dex") # With de-optimization
The previous part was about the OAT/VDEX formats and how to access to the underlying DEX. This part introduces the main API for the
The LIEF DEX module enables to get information about Java code such as String, classes name, dalvik bytecodes, …
As LIEF project is only focused on formats, there won’t be Dalvik disassembler in the DEX module.
The main API for a DEX file is in the
lief.DEX.File object. This object can be generated using:
>>> oat = lief.parse("SecSettings2.odex") >>> type(oat.dex_files) _pylief.DEX.File >>> vdex = lief.VDEX.parse("SecSettings2.odex") >>> type(vdex.dex_files) _pylief.DEX.File >>> dex = lief.DEX.parse("classes.dex") >>> type(dex) _pylief.DEX.File
Once created, we can access to the strings with the
>>> len(dex.strings) 23529 >>> for s in dex.strings: ... if http in s: ... print(s) https://analytics.mopub.com/i/jot/exchange_client_event https://app-measurement.com/a https://mobilecrashreporting.googleapis.com/v1/crashes:batchCreate?key= https://pagead2.googlesyndication.com/pagead/gen_204?id=gmob-apps https://plus.google.com/ https://ssl.google-analytics.com https://support.google.com/dfp_premium/answer/7160685#push https://www.google.com ...
for cls in dex.classes: if cls.source_filename: print(cls)
com.avast.android.sdk.antitheft.internal.protection.wipe.a - CalendarWiper.java - 3 Methods com.avast.android.account.internal.identity.a - AvastIdentityProvider.java - 17 Methods com.avast.android.account.internal.identity.d - FacebookIdentityProvider.java - 19 Methods com.avast.android.lib.wifiscanner.internal.b$a - WifiScannerComponentFactory.java - 1 Methods
In the DEX file format, there is a special attribute for classes that register the original source filename: source_file_idx. Some obfuscators mangle classes but keep this attribute! Since Java source filenames are associated with class names, we can easily recover the deobfuscated name uing:
for cls in dex.classes: if cls.source_filename: print(cls.pretty_name + ": ---> " + cls.source_filename)
com.avast.android.sdk.antitheft.internal.protection.wipe.a: ---> CalendarWiper.java com.avast.android.account.internal.identity.a ---> AvastIdentityProvider.java com.avast.android.account.internal.identity.d ---> FacebookIdentityProvider.java com.avast.android.lib.wifiscanner.internal.b$a ---> WifiScannerComponentFactory.java If we are interested with DEX methods, they are represented with the :class:`~lief.DEX.Method` object and we can access to the **raw** Dalvik bytecode through :attr:`lief.DEX.Method.bytecode`
ART is the name of the Android Runtime but it’s also a format! This format is used for optimization purpose by the Android’s framework.
As discussed previously, Android has its own implementation of the Java virtual based on the Dalvik bytecode.
This JVM is implemented in C++ and Java primitives (
java.lang.Object, etc) are mirrored with C++ objects:
When instantiating a new Java class, it creates a mirrored C++ object (Memory allocation, calling constructors, …) and the JVM handles a reference on this C++ object. To speed up the boot process and to avoid instantiation of well-known classes  at each boot, Android uses the ART format to store instances of C++ objects. To simplify, it can be seen as a heap dump of C++ objects.
In the same way as OAT and VDEX, the internal structures of this format change for each version of Android.
art = lief.ART.parse("boot.art") print(art.header)
Version: 46 Image Begin: 0x70000000 Image Size: 0x238ac8 Checksum: 0x997c0fb0 OAT File Begin: 0x70a5b000 OAT File End: 0x71272000 OAT Data Begin: 0x70a5c000 OAT Data End: 0x7126df70 Patch Delta: 0 Pointer Size: 8 Compile pic: true Number of sections: 10 Number of methods: 7 Boot Image Begin: 0 Boot Image Size: 0 Boot OAT Begin: 0 Boot OAT Size: 0 Storage Mode: UNCOMPRESSED Data Size: 0x2389f0
LIEF 0.9 is read-only on these formats but further versions should enable to modify them (Add methods, change names, patch checksum, …)
|||Usually the ones from the Android Framework|
|||They have other nices features like a disassembler, pseudo-code, … that are not cover in LIEF|
|||(1, 2) Dextra by Jonathan Levin: http://newandroidbook.com/tools/dextra.html|
|||(1, 2) vdexExtractor by Anestis Bechtsoudis: https://github.com/anestisb/vdexExtractor|
|||smali by JesusFreke: https://github.com/JesusFreke/smali/wiki/DeodexInstructions|