Repository files navigation

Simplify virtually executes an app to understand its behavior and then tries to optimize the code so that it behaves identically but is easier for a human to understand. Each optimization type is simple and generic, so it doesn't matter what the specific type of obfuscation is used.

The code on the left is a decompilation of an obfuscated app, and the code on the right has been deobfuscated.

There are three parts to the project: smalivm, simplify, and the demo app.

1. smalivm: Provides a virtual machine sandbox for executing Dalvik methods. After executing a method, it returns a graph containing all possible register and class values for every execution path. It works even if some values are unknown, such as file and network I/O. For example, anyif orswitch conditional with an unknown value results in both branches being taken.
2. simplify: Analyzes the execution graphs fromsmalivm and applies optimizations such as constant propagation, dead code removal, unreflection, and some peephole optimizations. These are fairly simple, but when applied together repeatedly, they'll decrypt strings, remove reflection, and greatly simplify code. It doesnot rename methods and classes.
3. demoapp: Contains simple, heavily commented examples for usingsmalivm in your own project. If you're building something that needs to execute Dalvik code, check it out.

usage: java -jar simplify.jar <input> [options]deobfuscates a dalvik executable -et,--exclude-types <pattern>   Exclude classes and methods which include REGEX, eg: "com/android", applied after include-types -h,--help                       Display this message -ie,--ignore-errors             Ignore errors while executing and optimizing methods. This may lead to unexpected behavior.    --include-support            Attempt to execute and optimize classes in Android support library packages, default: false -it,--include-types <pattern>   Limit execution to classes and methods which include REGEX, eg: ";->targetMethod\("    --max-address-visits <N>     Give up executing a method after visiting the same address N times, limits loops, default: 10000    --max-call-depth <N>         Do not call methods after reaching a call depth of N, limits recursion and long method chains, default: 50    --max-execution-time <N>     Give up executing a method after N seconds, default: 300    --max-method-visits <N>      Give up executing a method after executing N instructions in that method, default: 1000000    --max-passes <N>             Do not run optimizers on a method more than N times, default: 100 -o,--output <file>              Output simplified input to FILE    --output-api-level <LEVEL>   Set output DEX API compatibility to LEVEL, default: 15 -q,--quiet                      Be quiet    --remove-weak                Remove code even if there are weak side effects, default: true -v,--verbose <LEVEL>            Set verbosity to LEVEL, default: 0

Building requires theJava Development Kit 8 (JDK) to be installed.

Because this project contains submodules for Android frameworks, either clone with--recursive:

git clone --recursive https://github.com/CalebFenton/simplify.git

Or update submodules at any time with:

git submodule update --init --recursive

Then, to build a single jar which contains all dependencies:

./gradlew fatjar

The Simplify jar will be insimplify/build/libs/. You can test it's working by simplifying the providedobfuscated example app. Here's how you'd run it (you may need to changesimplify.jar):

java -jar simplify/build/libs/simplify.jar -it"org/cf/obfuscated" -et"MainActivity" simplify/obfuscated-app.apk

To understand what's getting deobfuscated, check outObfuscated App's README.

If Simplify fails, try these recommendations, in order:

1. Only target a few methods or classes by using-it option.
2. If failure is because of maximum visits exceeded, try using higher--max-address-visits,--max-call-depth, and--max-method-visits.
3. Try with-v or-v 2 and report the issue with the logs and a hash of the DEX or APK.
4. Try again, but do not break eye contact. Simplify can sense fear.

If building on Windows, and building fails with an error similar to:

Could not find tools.jar. Please check that C:\Program Files\Java\jre1.8.0_151 contains a valid JDK installation.

This means Gradle is unable to find a proper JDK path. Make sure the JDK is installed, set theJAVA_HOME environment variable to your JDK path, and make sure to close and re-open the command prompt you use to build.

Don't be shy. I think virtual execution and deobfuscation are fascinating problems. Anyone who's interested is automatically cool and contributions are welcome, even if it's just to fix a typo. Feel free to ask questions in the issues and submit pull requests.

Please include a link to the APK or DEX and the full command you're using. This makes it much easier to reproduce (and thusfix) your issue.

If you can't share the sample,please include the file hash (SHA1, SHA256, etc).

If an op places a value of a type which can be turned into a constant such as a string, number, or boolean, this optimization will replace that op with the constant. For example:

const-string v0, "VGVsbCBtZSBvZiB5b3VyIGhvbWV3b3JsZCwgVXN1bC4="invoke-static {v0}, Lmy/string/Decryptor;->decrypt(Ljava/lang/String;)Ljava/lang/String;# Decrypts to: "Tell me of your homeworld, Usul."move-result v0

In this example, an encrypted string is decrypted and placed intov0. Since strings are "constantizable", themove-result v0 can be replaced with aconst-string:

const-string v0, "VGVsbCBtZSBvZiB5b3VyIGhvbWV3b3JsZCwgVXN1bC4="invoke-static {v0}, Lmy/string/Decryptor;->decrypt(Ljava/lang/String;)Ljava/lang/String;const-string v0, "Tell me of your homeworld, Usul."

Code is dead if removing it cannot possibly alter the behavior of the app. The most obvious case is if the code is unreachable, e.g.if (false) { // dead }). If code is reachable, it may be considered dead if it doesn't affect any state outside of the method, i.e. it has noside effect. For example, code may not affect the return value for the method, alter any class variables, or perform any IO. This is a difficult to determine in static analysis. Luckily, smalivm doesn't have to be clever. It just stupidly executes everything it can and assumes there are side effects if it can't be sure. Consider the example from Constant Propagation:

const-string v0, "VGVsbCBtZSBvZiB5b3VyIGhvbWV3b3JsZCwgVXN1bC4="invoke-static {v0}, Lmy/string/Decryptor;->decrypt(Ljava/lang/String;)Ljava/lang/String;const-string v0, "Tell me of your homeworld, Usul."

In this code, theinvoke-static no longer affects the return value of the method and let's assume it doesn't do anything weird like write bytes to the file system or a network socket so it has no side effects. It can simply be removed.

const-string v0, "VGVsbCBtZSBvZiB5b3VyIGhvbWV3b3JsZCwgVXN1bC4="const-string v0, "Tell me of your homeworld, Usul."

Finally, the firstconst-string assigns a value to a register, but that value is never used, i.e. the assignment is dead. It can also be removed.

const-string v0, "Tell me of your homeworld, Usul."

Huzzah!

One major challenge with static analysis of Java is reflection. It's just not possible to know the arguments are for reflection methods without doing careful data flow analysis. There are smart, clever ways of doing this, but smalivm does it by just executing the code. When it finds a reflected method invocation such as:

invoke-virtual {v0, v1, v2}, Ljava/lang/reflect/Method;->invoke(Ljava/lang/Object;[Ljava/lang/Object;)Ljava/lang/Object;

It can know the values ofv0,v1, andv2. If it's sure what the values are, it can replace the call toMethod.invoke() with an actual non-reflected method invocation. The same applies for reflected field and class lookups.

For everything that doesn't fit cleanly into a particular category, there's peephole optimizations. This includes removing uselesscheck-cast ops, replacingLjava/lang/String;-><init> calls withconst-string, and so on.

.methodpublic statictest1()I.locals2new-instancev0,Ljava/lang/Integer;const/4v1,0x1invoke-direct {v0,v1},Ljava/lang/Integer;-><init>(I)Vinvoke-virtual {v0},Ljava/lang/Integer;->intValue()Imove-resultv0returnv0.end method

All this does isv0 = 1.

.methodpublic statictest1()I.locals2new-instancev0,Ljava/lang/Integer;const/4v1,0x1invoke-direct {v0,v1},Ljava/lang/Integer;-><init>(I)Vinvoke-virtual {v0},Ljava/lang/Integer;->intValue()Iconst/4v0,0x1returnv0.end method

Themove-result v0 is replaced withconst/4 v0, 0x1. This is because there is only one possible return value forintValue()I and the return type can be made a constant. The argumentsv0 andv1 are unambiguous and do not change. That is to say, there's a consensus of values for every possible execution path atintValue()I. Other types of values that can be turned into constants:

• numbers -const/4,const/16, etc.
• strings -const-string
• classes -const-class

.methodpublic statictest1()I.locals2const/4v0,0x1returnv0.end method

Because the code aboveconst/4 v0, 0x1 does not affect state outside of the method (no side-effects), it can be removed without changing behavior. If there was a method call that wrote something to the file system or network, it couldn't be removed because it affects state outside the method. Or iftest()I took a mutable argument, such as aLinkedList, any instructions that accessed it couldn't be considered dead.

Other examples of dead code:

• unreferenced assignments - assigning registers and not using them
• unreached / unreachable instructions -if (false) { dead_code(); }

This tool is available under a dual license: a commercial one suitable for closed source projects and a GPL license that can be used in open source software.

Depending on your needs, you must choose one of them and follow its policies. A detail of the policies and agreements for each license type are available in theLICENSE.COMMERCIAL andLICENSE.GPL files.

• Dalvik Virtual Execution with SmaliVM
• Guillot, Yoann, and Alexandre Gazet. "Automatic Binary Deobfuscation." Journal in Computer Virology 6.3 (2010): 261-76
• Unicorn - The ultimate CPU emulator
• Babak Yadegari, Saumya Debray. "Symbolic Execution of Obfuscated Code"
• Success stories:
  • Android Dynamic Class Loading with "AES/CFB/NoPadding" encryption. Take a peek before & after. Tool used: #simplify. #Android #obfuscation #classencryption #dex
  • Decrypting Malware String Encryption