KEVIN COOGAN, GEN LU, SAUMYA DEBRAY DEPARTMENT OF COMUPUTER SCIENCE UNIVERSITY OF ARIZONA 報告者：張逸文 Deobfuscation of Virtualization- Obfuscated Software.

ADLab

1

K E V I N C O O G A N , G E N L U , S A U M YA D E B R AY

D E PA RT M E N T O F C O M U P U T E R S C I E N C E U N I V E R S I T Y O F A R I Z O N A

報告者：張逸文

Deobfuscation of Virtualization-Obfuscated Software

ADLab

2

Outline

IntroductionDeobfuscationExperimental EvaluationRelated WorkConclusion

ADLab

3

Introduction（ 1/4）

Basic about Reverse Engineering Compilation

Decompilation

ADLab

4

Introduction（ 2/4）

Virtualization obfuscators VMProtect, Code Virtualizer

{

VIRTUALIZER_START

your code

VIRTUALIZER_END

}

ADLab

5

Introduction（ 3/4）

The virtualization-obfuscated programs are resistant to static and dynamic analysis techniques The executed code reveals only the structure and logic of the byte-

code interpreter Randomness VM

Outside-in approach Reverse engineer the VM interpreter Individual byte code instructions Recover the logic The structure of the interpreter meets certain requirements

ADLab

6

Introduction（ 4/4）

Programs interact with the system through system callsIdentifying instructions that interact with the systemNot recovering the original instructionsCapturing behavior of the codeGeneral, using in a wide range

ADLab

7

Deobfuscation

Static analysis v.s dynamic traceIdentifying instructions that are known to be part of the

original codeNo information about the specific structure of the

interpreter

ADLab

8

Deobfuscation

Overall approach：1. Tracing tool

Low level execution trace

2. Identifying system calls and their arguments database

3. Instruction trace Relevant instructions

4. Building a subtrace Relevant subtrace

9

Deobfuscation

ADLab

Value-based Dependence Analysis Not recovering the original code The process of deobfuscation must be semantics-preserving Identifying instructions that affect the values of the arguments to

system calls Slicing algorithms --- control-dependent Data dependencies Use-definition chains --- link instructions that use a variable to the

instruction that define it Problem：

10

Deobfuscation

ADLab

Value-based dependence

if( I defines a location l S) {

I is marked as relevant;

l is removed from S;

the set of locations used by I is added to S; } Problem： a pointer to a structure

I uses some locations l1, l2, … , ld

if ( I uses li P to define ld )

ld is added to P

if ( li access a memory location )

[li ] is added to M

ADLab

11

Deobfuscation

Relevant Conditional Control Flow Value-based dependence analysis doesn’t identify the associated

control flow instructions The occurring of conditional control flow IA-32 architecture setting the condition code flags in the eflags

register Not such simple!! Examining target address Equational Resoning System： translate each instruction in the

dynamic trace into an equivalent set of equations

ADLab

12

Deobfuscation

Equational Resoning System Identifies conditional dependencies The left hand side variables in an equation is numbered by the order of

its instruction appears The right hand side variables is numbered by the instruction that defined

it Example 1.

13

Deobfuscation

ADLab

Example 2. Example 3. Indirect jump

ADLab

14

Deobfuscation

Example 4. Used in VMProtect

Target20 = index1*4+0x10000

ADLab

15

Deobfuscation

16

Deobfuscation

ADLab

ADLab

17

Deobfuscation

Relevant Call-Return Control Flow Identifying functions： the behavior of calls and returns Knowing how them work allows one to use for other purposes Behavior of Function Calls and Returns

ADLab

18

Deobfuscation

registers

call 改成 push

無法解決

19

Deobfuscation

ADLab

Identification Approach Call： a code address is saved at the call site Return： the saved address is used for a control transfer at the return

point

ADLab

20

Deobfuscation

Relevant Dynamic Trace

21

Experimental Evaluation

ADLab

Experimental Methodology Compile original source code Generate an original dynamic trace Build an original subtrace Virtualization-obfuscation technique Generate an obfuscated dynamic trace Build a relevant subtrace of the obfuscated subtrace The obfuscated subtrace is matched to the original subtrace and

scores are produced The relevance score and obfuscation score are calculated

ADLab

22

Experimental Evaluation

VX Heavens website

ADLab

23

Related Work

Deobfuscation of code obfuscated via virtualization obfuscators Rolles, Sharif, Falliere

Programming language community Partial evaluation

ADLab

24

Conclusions

Virtualization-obfuscated programs are difficult to reverse engineer

We present a different approach to identifying the flow of values to system call instructions

ADLab

25

XD ~