SRI International

Reverse Engineering
Malware

Hassen Saidi
Computer Science Laboratory
SRI International
Hassen.Saidi@sri.com
The Growth of a Network
 The Growth of a Threat
The Growth of a Threat
Mass email campaign: Love letter, Melissa
Multiple vectors of infection, attacks against AV software,
Combined infection vectors, dangerous payloads: Code Red, Nimda

Email virus + social engineering: Xmas Exec

Large scale pandemics: Morris worm
Infected 10% of the Internet.

Self replicating Sophisticated engineering: Conficker

Program: Creeper Use of Crypto.
Social Networks/cell phone worms.
Stuxnet,…
 The Threats of the Connected Age
The Threats of the Connected Age

Malware incidents are

rising dramatically:
• increase of infection vectors
• increase in the complexity of
botnet structures
From Biology: Connected World Gives Viruses The Edge
“as human activity makes the world more connected,
natural selection will favor more virulent and
dangerous parasites."
 Motivation
• Malware landscape is diverse and constant evolving
– Large botnets
– Diverse propagation vectors, exploits, C&C
– Capabilities – backdoor, keylogging, rootkits,
– Logic bombs, time-bombs
– Diverse targets: desktops, mobile platforms, SCADA systems (Stuxnet)
• Malware is not about script-kiddies anymore, it’s real business.
Recent events indicate that it can be a powerful weapon in
cyber warfare.
• Manual reverse-engineering is close to impossible
– Need automated techniques to extract system logic, interactions and
side-effects, derive intent, and devise mitigating strategies.
 Outline
• Review of the workflow of binary program analysis
• Review of the challenges in binary program
analysis:
– Obfuscation Techniques
• Techniques for reverse engineering stripped
binaries:
– Systematic deobfuscation
• Examples of obfuscation: Conficker, Hydrac
(Google attack), Stuxnet, …
 Capturing Malware

• Honeynets: Capture malware that scans the

Internet for vulnerable targets
• Mining SPAM for attachments
• Mining SPAM for malicious URLs, and
capturing drive-by downloads
• AV heuristics
 Malware Binary Analysis

01001010100101010
10101010011010101
01001010100101010
10101010011010101
• What does the malware do
01001010100101010 • How does it do it
10101010011010101
01001010100101010
• identify triggers
10101010011010101 • What is the purpose of the
10101010011010101
01001010100101010
malware
10101010011010101 • is this an instance of a known
.exe threat or a new malware
• who is the author
Typically a stripped
•…
binary with no
debugging information.

In the case of malicious

Challenges:
code, it is often obfuscated • lack of automation
and packed • time-critical analysis
• labor intensive
Often has embedded suicide logic and • requires a human in the loop
anti-analysis logic
 Dynamic vs Static Malware Analysis

• Dynamic Analysis
– Techniques that profile actions of binary at
runtime
– More popular
• CWSandbox, TTAnalyze, multipath exploration
• Only provides partial ``effects-oriented profile’’ of
malware potential
• Static Analysis
– Can provide complementary insights
– Potential for more comprehensive assessment
 Malware Evasions and Obfuscations

• To defeat signature based detection schemes

– Polymorphism, metamorphism: started appearing in viruses of the 90’s
primarily to defeat AV tools
• To defeat Dynamic Malware Analysis
– Anti-debugging, anti-tracing, anti-memory dumping
– VMM detection, emulator detection
• To defeat Static Malware analysis
– Encryption (packing)
– API and control-flow obfuscations
– Anti-disassembly
• The main purpose of obfuscation is to slow down the security
community
 My Personal Philosophy

• Push the limits of static analysis as much as

possible.
• Rebuild the binary in its original form prior to
obfuscation.
• Recover a higher level description of the
malware binary that makes deriving the
purpose of the malware atteingnable: I want
to stare at C code as opposed to staring at
assembly code
 Malware Revere Engineering System Goals

• Desiderata for a Static Analysis Framework

– Unpack most of contemporary malware
– Handle most if not all packers
– Deobfuscate API references
– Automate identification of capabilities
– Provide feedback on unpacking success
– Simplify and annotate call graphs to illustrate interactions
between key logical blocks
– Enable decompilation of assemply code into a higher-level
language
– Identify key logical blocks (crypto for instance)
 Reverse Engineering Phases
 Unpacking phase: the image of a running malware sample is often considered
damaged:
- No known OEP. Imported APIs are invoked dynamically and the original import
table is destroyed. Arbitrary section names and r/w/e permissions.
 Disassembly phase:
- Identification of code and data segments
- Relies on the unpacker to capture all code and data segments. Our unpacking
approach guarantees that.
 Decompilation phase:
- Reconstruction of the code segment into a C-like higher level representation
- Relies on the disassembler to recognize function boundaries, targets of call
sites, imports, and OEP. Our API resolution guarantees that.
 Program understanding phase:
- Relies on the decompiler to produce readable C code, by recognizing the
compiler, calling conventions, stack frames manipulation, functions prologs and
epilogs, user-defined data structures. Our code rewrite and analysis guarantees
that.
Phase 1: Malware Unpacking

Unpacking
Example of Packed Code
 The Eureka Framework

• Novel unpacking technique based on coarse

grained execution tracing
• Heuristic-based and statistic-based upacking
• Implements several techniques to handle
obfucated API references
• Multiple metrics to evaluate unpack success
• Annotated call graphs provide bird’s eye view
of system interaction
The Eureka Workflow
 Coarse-grained Execution Monitoring

• Generalized unpacking principle

– Execute binary till it has sufficiently revealed itself
– Dump the process execution image for static
analysis
• Monitoring exection progress
– Eureka employs a Windows driver that hooks to
SSDT (System Service Dispatch Table)
– Callback invoked on each NTDLL system call
– Filtering based on malware process pid
 Heuristic-based Unpacking

• How do you determine when to dump?

– Heuristic #1: Dump as late as possible. NtTerminateProcess
– Heuristic #2: Dump when your program generates errors.
NtRaiseHardError
– Heuristic #3: Dump when program forks a child process.
NtCreateProcess
• Issues
– Weak adversarial model, too simple to evade…
– Doesn’t work well for package non-malware programs
 Statistics-based Unpacking

• Observations
– Statistical properties of packed executable differ
from unpacked exectuable
– As malware executes code-to-data ratio increases
• Complications
– Code and data sections are interleaved in PE
executables
– Data directories(import tables) look similar to data
but are often found in code sections
– Properties of data sections vary with packers
 Statistics-based Unpacking (2)

• Our Approach
– Model statistical properties of unpacked code
• Estimating unpacked code
– N-gram analysis to look for frequent instructions
– We use bi-grams (2-grams) because x-86 opcodes are 1 or 2
bytes
– Extract subroutine code from 9 benign executables
– FF 15 (call), FF 75 (push), E8 _ _ _ ff (call), E8 _ _ _ 00 (call)
Evaluation (ASPack)
Evaluation (MoleBox)
Evaluation (Armadillo)
 Systematic Approach to Code Deobfuscation:
Unpacking

• Automatic Unpacking: involves running the malware and capturing

its memory image.
• Monitoring the execution of the malware is an intrusive process and
is often detected using anti-tracing and anti-debugging techniques
embedded in the malware.
• Our multi-strategy approach consists of minimal monitoring and
capturing the process image at key events:
– ExitProcess
– Byte bigram monitoring: call, push instructions for instance
– Number of seconds elapsed
– Run the malware without monitoring and suspend its execution and perform memory
inspection
• In practice, we always manage to get a dump (memory snapshot) of
the running process: no OEP and no Import table
 Phase 2: Disassembly
• The disassembler reads the PE data structure in order to:
1. Determine the different sections of the file and separate code from data and
identifies resource information such as import tables
– The disassembler relies on the PE data structure (could be corrupt)
– The disassembler translates into code, any referenced address from
known code location
2. Translate code segments into assembly language
– The disassembler relies on the hardware instruction set documentation
3. Interpret data according to identified types
1. A data referenced by code can be of any type: integer, string, struct, etc.
Integer:
0x0040F45C dword_40F45C dd 0E06D7363h, 1, 2 dup(0) ; DATA XREF: 408C98
String:
0x0040F45C unk_40F45C db 63h ; c ; DATA XREF: sub_408C98
0x0040F45D db 73h ; s
0x0040F45E db 6Dh ; m
0x0040F45F db 0E0h ; a
 IDA Pro Disassembler
• http://www.hex-rays.com/idapro/
– It supports a variety of executable formats for different processors and
operating systems. It also can be used as a debugger for Windows PE, Mac
OS X, and Linux ELF executables.

– IDA performs a large degree of automatic code analysis to a certain extent,

leveraging cross-references between code sections, knowledge of
parameters of API calls, limited dataflow analysis, and recognition of
standard libraries.
• Hashes of known statically linked libraries are compared to hashes of identified
subroutines in the code

– Provides scripting languages to interact with the system to improve the

analysis.

– Support plug-ins: The IDA decompiler is the most impressive plug-in.

PE Execution

1. Read the Portable Executable

(PE) file data structure and
maps the file into memory

2. Load import modules

1. Start execution at entry point

2. Runtime unpacking

3. Jump to OEP
 Phase 3: Fixing the Disassembled Code

• Unpacked & disassembled code does not have

an OEP.
• Import tables are rebuilt dynamically and
there are no static references to dynamically
loaded libraries
• Header information is not reliable
• Data is not typed
Parsing the PE executable format
 Challenges in Binary Code Disassembly

• Disassembly is not an exact science: On CISC platforms with

variable-width instructions, or in the presence of self-
modifying code, it is possible for a single program to have two
or more reasonable disassemblies. Determining which
instructions would actually be encountered during a run of
the program reduces to the proven-unsolvable halting
problem.

• Bad disassembly because of variable length instructions

• Jumps into middle of instructions
• No reachability analysis: Unreachable code can hide data.
 Examples of Disassembly problems (The Storm Worm)

Data hidden as code:

ArcadeWorld.exe:0042BDB0 mov eax, 0
ArcadeWorld.exe:0042BDB5 test eax, eax
ArcadeWorld.exe:0042BDB7 jnz short loc_42BDD8; unreachable code

ArcadeWorld.exe:0042BDD8 loc_42BDD8: ; CODE

XREF: 0042BDB7 0042BDD8 8D 95 42 1C 40 00 8D 85 20 B2 40 00 50 68 20 8A ìòB.@.ìà ¦@.Ph è
ArcadeWorld.exe:0042BDD8 lea edx, [ebp+401C42h] 0042BDE8 00 00 52 E8 8D 01 00 00 60 8B BD 67 C0 40 00 03 ..RFì...`ï+g+@..
ArcadeWorld.exe:0042BDDE lea eax, [ebp+40B220h] 0042BDF8 BD 3F C0 40 00 8D B5 BF BF 40 00 B9 74 00 00 00 +?+@.ì¦++@.¦t...
ArcadeWorld.exe:0042BDE4 push eax 0042BE08 68 00 20 00 00 57 E8 26 FF FF FF F3 A4 8D 85 AF h. ..WF& =ñìà»
ArcadeWorld.exe:0042BDE5 push 8A20h 0042BE18 BF 40 00 8B 9D 3F C0 40 00 FF B5 43 C0 40 00 FF +@.ï¥?+@. ¦C+@.
ArcadeWorld.exe:0042BDEA push edx 0042BE28 B5 37 C0 40 00 6A 01 50 53 E8 9E 00 00 00 FF B5 ¦7+@.j.PSFP... ¦
ArcadeWorld.exe:0042BDEB call near ptr unk_42BF7D 0042BE38 6B C0 40 00 FF B5 3F C0 40 00 E8 DE FD FF FF 8B k+@. ¦?+@.F¦² ï
ArcadeWorld.exe:0042BDF0 pusha 0042BE48 85 18 B2 40 00 85 C0 74 1C FF B5 18 B2 40 00 FF à.¦@.à+t. ¦.¦@.
ArcadeWorld.exe:0042BDF1 mov edi, [ebp+40C067h] 0042BE58 B5 57 C0 40 00 FF B5 3F C0 40 00 E8 E7 10 00 00 ¦W+@. ¦?+@.Ft...
ArcadeWorld.exe:0042BDF7 add edi, [ebp+40C03Fh] 0042BE68 E8 3F FE FF FF 53 E8 E5 04 00 00 E8 14 00 00 00 F?¦ SFs...F....
ArcadeWorld.exe:0042BDFD lea esi, [ebp+40BFBFh] 0042BE78 5C 64 6C 6C 63 61 63 68 65 5C 74 63 70 69 70 2E \dllcache\tcpip.
0042BE88 73 79 73 00 E8 16 0A 00 00 E8 13 00 00 00 5C 64 sys.F....F....\d
 API Resolution

• User-level malware programs require system

calls to perform malicious actions
• Use Win32 API to access user level libraries
• Obufscations impede malware analysis using
IDA Pro or OllyDbg
– Packers use non-standard linking and loading of
dlls
– Obfuscated API resolution
 Standard API Resolution
Imports in IAT identified by IDA by looking at Import Table
 Handling Thunks
• Identify subroutines with a JMP instruction only
– Treat any calls to these subs as an API call
IsDebuggerPresent
 Leveraging Standard API Address Loading
==================================================
Function Name : ADSICloseDSObject
Address : 0x76e30826
Relative Address : 0x00020826
Ordinal : 142 (0x8e)
Filename : adsldpc.dll
Full Path : c:\WINDOWS\system32\adsldpc.dll
Type : Exported Function
==================================================

==================================================
Function Name : ADSICloseSearchHandle
Address : 0x76e3050a
Relative Address : 0x0002050a
Ordinal : 143 (0x8f)
Filename : adsldpc.dll
Full Path : c:\WINDOWS\system32\adsldpc.dll
Type : Exported Function
==================================================

==================================================
Function Name : ADSICreateDSObject
Address : 0x76e30447
Relative Address : 0x00020447
Ordinal : 144 (0x90)
Filename : adsldpc.dll
Full Path : c:\WINDOWS\system32\adsldpc.dll
Type : Exported Function
==================================================
 Using Dataflow Analysis
• Identify register based indirect calls
GetEnvironmentStringW

def

use
 Handling Dynamic Pointer Updates
• Identify register based indirect calls

A def to dword_41e308 is found

Look for probable call to
GetProcAddress earlier dword_41e304 has no static
Call to GetProcAddress value to look up API

def

use
 Leveraging Standard API Address Loading is not enough
==================================================
Function Name : ADSICloseDSObject
Address : 0x76e30826
Relative Address : 0x00020826
Ordinal : 142 (0x8e)
Filename
Full Path
: adsldpc.dll
: c:\WINDOWS\system32\adsldpc.dll
There are many indirect ways to load
Type : Exported Function
==================================================
And call a Windows API:
• access to list of loaded DLLs
==================================================
Function Name : ADSICloseSearchHandle • access to a loaded DLL and use of
Address : 0x76e3050a
Relative Address : 0x0002050a GetModulHandle() + offset
Ordinal
Filename
: 143 (0x8f)
: adsldpc.dll
•…
Full Path : c:\WINDOWS\system32\adsldpc.dll
Type : Exported Function
==================================================

==================================================
Function Name : ADSICreateDSObject
Address : 0x76e30447
Relative Address : 0x00020447
Ordinal : 144 (0x90)
Filename : adsldpc.dll
Full Path : c:\WINDOWS\system32\adsldpc.dll
Type : Exported Function
==================================================
 Consequence of Failure to Identify APIs

...
.text:004011A7 push offset unk_40A2DC ; arg 1 Name of a library
.text:004011AC xor ebx, ebx
.text:004011AE call dword ptr unk_40A0E4 .data:0040A0E4 00000000 Load library call (LoadLibrary)
.text:004011B4 mov edi, eax
.text:004011B6 cmp edi, ebx
.text:004011B8 jz short loc_401211
.text:004011BA push esi
.text:004011BB mov esi, dword ptr unk_40A0E8 Name of the library function
.text:004011C1 push offset unk_40A2C4 ; arg 2
.text:004011C6 push edi ; arg 1
Name of the library
.text:004011C7 call esi ; unk_40A0E8 .data:0040A0E8 00000000 API call to get the address
.text:004011C9 push offset unk_40A2AC Of the loaded library function
.text:004011CE push edi (GetProcAddress)
.text:004011CF mov dword_433480, eax
...

...
.text:00401132 lea eax, [ebp+var_4]
.text:00401135 push eax
.text:00401136 push ebx
.text:00401137 push 0
.text:00401139 mov [ebp+var_4], esi
.text:0040113C call dword_433480 library function call
.text:00401142 test eax, eax
…
 Failure to Perform Control Flow Analysis
• CreateThread

.text:009A3A4C push eax

.text:009A3A4D xor eax, eax
.text:009A3A4F push eax
Location of the start address of a thread
.text:009A3A50 push eax
.text:009A3A51 push offset dword_9A3939
.text:009A3A56 push eax
Call to CreateThread
.text:009A3A57 push eax
.text:009A3A58 call [ebx]

.data:009A3939 xxxxxxx

• Starting Services
• Thread synchronization
• Critical sections
• Callback functions
 Advanced API Resolution

• There are many ways in which a library or API

can be invoked.
• There are many ways an API call can be
obfuscated
• But there is one invariant associated to each
API and library: its signature
– i.e; number of arguments, type of arguments, and
type of return value if any.
 Advanced API Resolution: Type Inference for binary program
analysis
• Use type inference as a single solution to solve three fundamental problems:
– Identifying API and function calls (call and jump targets)
– Building a precise CFG
– Recovering user-defined types for proper decompilation

• For Windows Executable files:

– Integers: object handles, addresses, IP address, ports, etc
– Strings: file names, service names, etc
– Structures: sockaddr

struct sockaddr_in {
short sin_family;
u_short sin_port;
struct in_addr sin_addr;
char sin_zero[8];
};
 Type propagation and matching
• Type propagation using dataflow analysis
• Propagation of return values and arguments of functions

sub_403649 proc near There is only one API that has

...
.text:00403649 arg_0 = dword ptr 8
7 arguments such that the seventh
.text:00403649 arg_4 = dword ptr 0Ch and third and first one can be
.text:00403649 arg_8 = dword ptr 10h pointers and all others are not.
.text:00403649 arg_C = dword ptr 14h
...
.text:00403668 xor ebx, ebx
...
.text:00403704 mov esi, ds:dword_40A06C
HANDLE WINAPI CreateFile(
.text:0040370A push ebx ; arg 7 type(f,7) = type (ebx)
.text:0040370B mov edi, 80h __in LPCTSTR lpFileName,

.text:00403710 push edi ; arg 6 type(f,6) = type (edi) __in DWORD dwDesiredAccess,
.text:00403711 push 4 ; arg 5 type(f,5) = union(int,char) __in DWORD dwShareMode,
.text:00403713 push ebx ; arg 4 type(f,4) = type (ebx) __in LPSECURITY_ATTRIBUTES lpSecurityAttributes,
.text:00403714 push 2 ; arg 3 type(f,3) = union(int,char) __in DWORD dwCreationDisposition,
.text:00403716 push 2 ; arg 2 type(f,2) = union(int,char)
__in DWORD dwFlagsAndAttributes,
.text:00403718 push [ebp+arg_4] ; arg 1 type(f,1) = type([ebp+arg_4])
__in HANDLE hTemplateFile
.text:0040371B mov [ebp+var_18], ebx
);
.text:0040371E call esi ; dword_40A06C type(ret(f)) = type(eax)
...
 Advantages of type Inference Analysis
• Programmers data structures and types are going to be based on
known data structures and types provided by the libraries

• Identifying API calls and type information help capture better the
semantics of the program execution

• Not restricted to Windows but require knowledge of the libraries

and their documentation

• Can deal with some of the widely used obfuscation techniques

– Import table obfuscation
– Code rewrite: code rewrite preserves the types!
 Phase 3: Rebuilding the unpacked executable

• From a damaged dumped image of a running malware to a PE

executable:
– Knowing all APIs allows us to identify the OEP.
– Semantic approach: ExitProcess, CreateMutex, GetCommandLine,
GetModulaHandle, etc are close to OEP.There are about 20 APIs that are
often called at the beginning of the execution of the code.
– Structural approach: find sources of call graphs in the binary
– Rebuilding in import table with all references to identified APIs
• The disassembly of the reconstructed PE is often of better
quality than the disassembly of the dumped process image
– The new PE code bypasses the unpacking routine embedded in the
packed code
– The new PE contains the original code
 Phase 4: Decompilation

• Identifies local variables

• Identifies arguments: registers, stack, or any
combination
• Identifies global variables
• Identify calling conventions
• Identifies common idioms and compiler features
• Eliminates the use of registers as intermediate
variables
• Identifies control structures
Decompilation Depends on Previous Analysis Phases
 Malware Obfuscation Effect on Decompilation

• While packing is the most used obfuscation

technique, it is often combined with other advanced
forms of obfuscation that make decompilation often
impossible:
• Call obfuscation in general and API obfuscation
in particular
• Binary Rewrite to create semantically
equivalent code with vastly different structure
• Chuncking or “code spaghettisation”
•…
 Analysis Phases
Ideally Reality
Source Code Malware

Compiler Unpacking

Executable code Non-executable

code
Disassembly & Disassembly & Undo
Analysis Analysis Obfuscation
Assembly code Obfuscated
Assembly code
assembly
Decompilation Decompilation Decompilation

Legitimate C/C++ A mess Legitimate C/C++

that a compiler
would generate
 Example of Binary Rewrite
.text:009B327C OBFUSCATED_VERSION_OF_is_private_subnet proc near
.text:009B327C mov ecx, eax .text:009A1311 is_private_subnet proc near
.text:009B327E and ecx, 0FFFFh .text:009A1311 mov ecx, eax
.text:009B3284 cmp ecx, 0A8C0h .text:009A1313 and ecx, 0FFFFh
.text:009B328A jz loc_9B4264 .text:009A1319 cmp ecx, 0A8C0h
.text:009B3290 jmp off_9BAAA5 .text:009A131F jz loc_9A1340
/* .text:009BAAA5 off_9BAAA5 dd offset loc_9AB10C */ .text:009A1325 cmp al, 0Ah
.text:009B3290 OBFUSCATED_VERSION_OF_is_private_subnet endp .text:009A1327 jz loc_9A1340
.text:009A132D and eax, 0F0FFh
. .text:009A1332 cmp eax, 10ACh
.text:009B4264 loc_9B4264: .text:009A1337 jz loc_9A1340
.text:009B4264 .text:009A133D xor eax, eax
.text:009B4264 mov eax, 1 .text:009A133F retn
.text:009B4269 retn .text:009A1340
.text:009A1340 loc_9A1340:
. .text:009A1340 mov eax, 1
.text:009AB10C cmp al, 0Ah .text:009A1345 retn
.text:009AB10E jz loc_9B4264 .text:009A1345 is_private_subnet endp
.text:009AB114 jmp off_9BA137
/* .text:009BA137 off_9BA137 dd offset loc_9B2CE4 */
.
.text:009B2CE4 and eax, 0F0FFh
.text:009B2CE9 cmp eax, 10ACh
.text:009B2CEE jz loc_9B4264
.text:009B2CF4 jmp off_9BA3CC
/* .text:009BA3CC off_9BA3CC dd offset loc_9ABA24 */
.
.text:009ABA24 xor eax, eax bool __usercall is_private_subnet (unsigned __int16 a1) {
.text:009ABA26 retn return a1 == 43200 || a1 == 10 || (a1 & 0xF0FF) == 4268;
}
int __usercall OBFUSCATED_VERSION_OF_is_private_subnet<eax>(unsigned __int16 a1 <ax>)
{
int result; // eax@2

if ( a1 == 43200 ) result = 1;
else result = off_9BAAA5();
return result;
}
 Systematic Approach to Code Deobfuscation:
Binary Rewriting

• Dechunking: The control flow of Conficker's P2P module has been significantly
obfuscated to hinder its disassembly and decompilation. Specifically, the contents of code blocks
from each subroutine have been extracted and relocated throughout different portions of the
executable. These different blocks (or chunks) are then referenced through unconditional and
conditional jump instructions. In effect, the logical control flow of the P2P module has been
obscured (spaghetti-code) to a degree that the module cannot be decompiled into coherent C-like
code, which typically drives more in-depth and accurate code interpretation. Move all blocks to a
contiguous memory block.

• Normalize x86 instructions: push followed by a pop is a

mov

• Normalize calling convention: cdecl, fastcall, stdcall,

instead of user-defined.
 Conficker and Hydrac Dechunking

• Identify all chunks in a function and rewrite the function

• Applied to all Conficker C P2P Protocol subroutines
• Unlike the Conficker P2P logic, Hydraq did not exhibit the same
level of obfuscation. It did, however, share some obfuscation
features with Conficker. The functions of the Hydraq binary
have been subjected to chunking, which renders decompilation
difficult. We applied our transformations to automatically
generate the C-like code for each subroutine and build a
complete CFG of the binary. The IDA disassembler identified
185 subroutines in the binary prior to our analysis. After
running the dechunking transformation, only 141 subroutine
remained and were decompiled.
 Purpose of code obfuscation
• While packing is often used to reduce the size of binaries
and to create polymorphic malware samples, the more
advanced obfuscation techniques are designed to slow
down reverse engineering efforts and to prevent:
– the identification of API calls: identify the basic building
blocks of the malware
– the control-flow reconstruction of the malware: follow
and reconstruct the logic flow
– static analysis: determine the full functionality, triggers,
hidden logic, time bombs, etc.
– timely reverse engineering and mitigation of the threat
 Why Code Obfuscation is not Easy

• Malware authors can design binary code that is extremely difficult

to analyze. Using advanced programming languages knowledge, it is
possible to create such code.
• Malware authors do not feel the need to always obfuscated their
code. Can easily defeat signature-based detection. Overwhelm
analysts and tools with large numbers of samples.
• Malware code should be able to run in a reliable manner.
Obfuscation should not compromise this important requirement
and should maintain the reliability of the initial code. This requires a
proof or guarantee of some sort.
• Malware deobfuscation is therefore a more attainable than you
might think. Systematic obfuscation informs systematic
deobfuscation.
 Our Approach

• Because obfuscation is introduced in a rather systematic way,

there is a hope that it can be dealt with in an automated way.

• Systematically identifying an obfuscation step and undoing its

effect.

• Focus on generic approaches as opposed to packer/obfuscator

specifics

• Focus on metrics that allow us to assess the effect and success

of our deobfuscation strategies
 Example: Static Analysis of Conficker
• Conficker appeared on November 20th, 2008
• Infected millions of machines worldwide
• Millions of machines still infected despite an extensive news coverage
about the threat
• Four versions have been released: A, B, B++, and C
• It is a sophisticated piece of malicious code created by professionals who
have extensive knowledge about networking, cryptography, system and
network programming, and security
• Managed to defeat the security community in stopping its progression by
using strong crypto, code obfuscation, aggressive propagation strategies,
and constantly monitoring the security community actions
• Dynamic analysis provided a limited understanding of the threat:
• Identification of what appears to be a P2P protocol
• Identification of ports opened by the malware
• Deobfuscation and static analysis were the only techniques that were able
to uncover the full capability of the malware.

SRI Technical Report (http://mtc.sri.com/Conficker)

 Example: Static Analysis of Conficker

Static Analysis of Conficker Code:

– Domain generation algorithm: provided a list of daily domains to be blocked
– Quarter of the Internet scanned: Understand what part of the Internet was
targeted for scanning and what infections were due to USB ports and mobile
devices
– List of disabled security products: detection
– Ukrainian keyboard avoidance: Geo-location database poisoning
– Use of MD6 and related crypto algorithms: Attribution
– DNS APIs patching to disable list of websites (including SRI!): detection
– Distribute a number of modified versions of the binary
– TCP and UDP ports based on the IP address of the infected machine:
detection
 Deobfuscation of the Conficker C P2P protocol

• Heavily obfuscated protocol code

• 88 APIs obfuscated
• Use of chunking lead to poor decompilation
• Benefits of the deobfuscation
– P2P Protocol description: protocol understanding and P2P structure
– Peer selection algorithm: proved the peer poisoning approach useless
– Possibility to hot patch code without DGA updates: proved C&C domains
obsolete

The P2P protocol was not just a mechanism for distributing PE executable
files but also digitally signed sets of x86 instructions that are executed in a
separate thread and take as argument the IP address of the sender. This
would provide a hot patch mechanism for all data manipulated by Conficker:
list of peers, encryption/decryption keys, the Conficker code it self, etc.
 Stuxnet: Keeping it “relatively” simple

• Stuxnet does not use advanced binary obfuscation techniques.

• The analysis of the code is challenging nevertheless
• Stuxnet Code Characteristics:
– Use of C++
– Use of C++ exception handling
– Use of C++ classes
– Use of simple data encoding (encryption)
– Use of C structures for all data passed to the main subroutines:
• Over 40 user-defined structures
• Not recognized by disassemblers and decompilers
 Phase 5: Program Understanding

• Need to identify higher-level concepts from

the deobfuscated code
• Need to interpret the code into a higher-level
malware objective
• Need to indentify particular features: crypto:
– Functions that use crypto-related opcode, loops,
etc
– Known constants in crypto algorithms
Finding Known and Unknown Crypto
 Conclusions
• It is always desirable to recover from the malware a
description that is as close as possible to the original
code produced by the authors.
• It is often possible to do that in practice
• It is often the only way to really determine the full
capability of the malware
• The benefits are important when it comes to high-
profile targets
• Easily integrated in common analysis tools:
disassembler (IDA), Decompilers.
 Daily Malware Capture and Analysis

• http://mtc.sri.com/