HUNTING FOR METAMORPHIC ENGINES Mark Stamp & Wing Wong September 13, 2006

Outline I. Metamorphic software Both good and evil uses II. Metamorphic virus construction kits III. How effective are metamorphic engines? How to compare two pieces of code? Similarity within and between virus families Similarity to non-viral code IV. Can we detect metamorphic viruses? Commercial virus scanners Hidden Markov models (HMMs) Similarity index V. Conclusion

PART I Metamorphic Software

What is Metamorphic Software?  Software is metamorphic provided All copies do the same thing Internal structure of copies differs  Today almost all software is cloned  “Good” metamorphic software… Mitigate buffer overflow attacks  “Bad” metamorphic software… Avoid virus/worm signature detection

Metamorphic Software for Good?  Suppose program has a buffer overflow  If we clone the program One attack breaks every copy Break once, break everywhere (BOBE)  If instead, we have metamorphic copies Each copy still has a buffer overflow One attack does not work against every copy BOBE-resistant Analogous to genetic diversity in biology  A little metamorphism does a lot of good!

Metamorphic Software for Evil?  Cloned virus/worm can be detected Common signature on every copy Detect once, detect everywhere (DODE?)  If instead virus/worm is metamorphic Each copy has different signature Same detection does not work against every copy Provides DODE-resistance Analogous to genetic diversity in biology  But, effective use of metamorphism here is tricky!

Crypto Analogy  In information security, almost everything that consistently works is either Crypto, or Has a crypto analogy…  Consider WWII ciphers  German Enigma Broken by Polish and British cryptanalysts Design was (mostly) known to cryptanalysts  Japanese Purple Broken by American cryptanalysts Design was (mostly) unknown to cryptanalysts

Crypto Analogy  Cryptanalysis  break a (known) cipher  Diagnosis  determine how an unknown cipher works (from ciphertext)  Which was the greater achievement, breaking Enigma or Purple? Cryptanalysis of Enigma was harder Diagnosis of Purple was harder  Can make a reasonable case for either…

Crypto Analogy  What does this have to do with metamorphic software?  Suppose we (the good guys) generate metamorphic copies of our software  Bad guys can attack individual copies  Can bad guys attack all copies? Bad guys can try to diagnose our metamorphic generator

Crypto Analogy  How to diagnose metamorphic generator (from exe’s)? Reverse engineer many copies, look at differences, etc., etc. Lots of work  Diagnosis problem is hard If good guys can force bad guys to solve a diagnosis problem, the good guys win  Security by obscurity? Violates (spirit of) Kerckhoffs’ Principle? Yes, but still may be valuable in the real world

Crypto Analogy  What about case where bad guys write metamorphic code? Metamorphic viruses, for example  Do good guys need to solve diagnosis problem? If so, good guys are in trouble  Not if good guys “only” need to detect the metamorphic code (not diagnose)  Not claiming the good guys job is easy  Just claiming that there is hope…

Virus Evolution  Viruses first appeared in the 1980s Fred Cohen  Viruses must avoid signature detection Virus can alter its “appearance”  Techniques employed encryption polymorphic metamorphic

Virus Evolution - Encryption  Virus consists of decrypting module (decryptor) encrypted virus body  Different encryption key different virus body signature  Weakness decryptor can be detected

Virus Evolution – Polymorphism  Try to hide signature of decryptor  Can use code emulator to decrypt putative virus dynamically  Decrypted virus body is constant Signature detection is possible

Virus Evolution – Metamorphism  Change virus body  Mutation techniques: permutation of subroutines insertion of garbage/jump instructions substitution of instructions

PART II Virus Construction Kits

Virus Construction Kits – PS-MPC  According to Peter Szor: “… PS-MPC [Phalcon/Skism Mass- Produced Code generator] uses a generator that effectively works as a code-morphing engine…… the viruses that PS-MPC generates are not [only] polymorphic, but their decryption routines and structures change in variants…”

Virus Construction Kits – G2  From the documentation of G2 ( Second Generation virus generator ): “… different viruses may be generated from identical configuration files…”

Virus Construction Kits - NGVCK  From the documentation for NGVCK ( Next Generation Virus Creation Kit ): “… all created viruses are completely different in structure and opcode…… impossible to catch all variants with one or more scanstrings.…… nearly 100% variability of the entire code”  Oh, really?

PART III How Effective Are Metamorphic Engines?

How We Compare Two Pieces of Code

Virus Families – Test Data  Four generators, 45 viruses 20 viruses by NGVCK 10 viruses by G2 10 viruses by VCL32 5 viruses by MPCGEN  20 normal utility programs from the Cygwin bin directory

Similarity within Virus Families – Results

NGVCK Similarity to Virus Families  NGVCK versus other viruses 0% similar to G2 and MPCGEN viruses 0 – 5.5% similar to VCL32 viruses (43 out of 100 comparisons have score > 0) 0 – 1.2% similar to normal files (only 8 out of 400 comparisons have score > 0)

NGVCK Metamorphism/Similarity  NGVCK By far the highest degree of metamorphism of any kit tested Virtually no similarity to other viruses or normal programs Undetectable???

PART IV Can Metamorphic Viruses Be Detected?

Commercial Virus Scanners  Tested three virus scanners eTrust version avast! antivirus version 4.7 AVG Anti-Virus version 7.1  Each scanned 37 files 10 NGVCK viruses 10 G2 viruses 10 VCL32 viruses 7 MPCGEN viruses

Commercial Virus Scanners  Results eTrust and avast! detected 17 (G2 and MPCGEN) AVG detected 27 viruses (G2, MPCGEN and VCL32) none of NGVCK viruses detected by the scanners tested

Hidden Markov Models (HMMs)  state machines  transitions between states have fixed probabilities  each state has a probability distribution for observing a set of observation symbols  can “train” an HMM to represent a set of data (in the form of observation sequences)  states = features of the input data  transition and the observation probabilities = statistical properties of features

HMM Example – the Occasionally Dishonest Casino

 2 states: fair/loaded  The switch between dice is a Markov process  Outcomes of a roll have different probabilities in each state  If we can only see a sequence of rolls, the state sequence is hidden  want to understand the underlying Markov process from the observations

HMMs – the Three Problems 1. Find the likelihood of seeing an observation sequence O given a model, i.e. P(O | ) 2. Find an optimal state sequence that could have generated a sequence O 3. Find the model parameters given a sequence O, i.e. find transition and observation probabilities that maximize the probability of observing O  There exist efficient algorithms to solve the three problems

HMM Application – Determining the Properties of English Text  Given: a large quantity of written English text  Input: a long sequence of observations consisting of 27 symbols (the 26 lower- case letters and the word space)  Train a model to find the most probable parameters (i.e., solve Problem 3)  Use trained model to score any unknown sequence of letters (and spaces) to determine whether it corresponds to English text. (i.e., solve Problem 1)

HMM Application – Initial and Final Observation Probability Distributions

HMM Application - Results  Observation probabilities converged, each letter belongs to one of the two hidden states  The two states correspond to consonants and vowels  Note: no a priori assumption was made HMM effectively recovered the statistically significant feature inherent in English

HMM Application - Results  Probabilities can be sensibly interpreted for up to n = 12 hidden states  Trained model could be used to detect English text, even if the text is “disguised” by, say, a simple substitution cipher or similar transformation

Virus Detection with HMMs  Use hidden Markov models (HMMs) to represent statistical properties of a set of metamorphic virus variants Train the model on family of metamorphic viruses Use trained model to determine whether a given program is similar to the viruses the HMM represents

Virus Detection with HMMs  A trained HMM maximizes the probabilities of observing the training sequence assigns high probabilities to sequences similar to the training sequence represents the “average” behavior if trained on multiple sequences represents an entire virus family, as opposed to individual viruses

Virus Detection with HMMs – Data  Data set 200 NGVCK viruses (160 for training, 40 for testing)  Comparison set 40 normal exes from Cygwin 25 other “non-family” viruses (G2, MPCGEN and VCL32)  25 HMM models generated and tested

Virus Detection with HMMs – Methodology

Virus Detection with HMMs – Results

 Detect some other viruses “for free”

Virus Detection with HMMs  Summary of experimental results All normal programs distinguished VCL32 viruses had scores close to NGVCK family viruses With proper threshold, 17 HMM models had 100% detection rate and 10 models had 0% false positive rate No significant difference in performance between HMMs with 3 or more hidden states

Virus Detection with HMMs – Trained Models  Converged probabilities in HMM matrices may give insight into the features of the represented viruses  We observe opcodes grouped into “hidden” states most opcodes in one state only  What does this mean? We are not sure…

HMMs – The Trained Models

Detection via Similarity Index  Straightforward similarity index can be used as detector To determine whether a program belongs to the NGVCK virus family, compare it to any randomly chosen NGVCK virus NGVCK similarity to non-NGVCK code is small Can use this fact to detect metamorphic NGVCK variants

Detection via Similarity Index

 Experiment compare 105 programs to one selected NGVCK virus  Results 100% detection, 0% false positive  Does not depend on specific NGVCK virus selected

PART V Conclusion

 Metamorphic generators vary a lot NGVCK has highest metamorphism (10% similarity on average) Other generators far less effective (60% similarity on average) Normal files 35% similar, on average  But, NGVCK viruses can be detected! NGVCK viruses too different from other viruses and normal programs

Conclusion  NGVCK viruses not detected by commercial scanners we tested  Hidden Markov model (HMM) detects NGVCK (and other) viruses with high accuracy  NGVCK viruses also detectable by similarity index

Conclusion  All metamorphic viruses tested were detectable because High similarity within family and/or Too different from normal programs  Effective use of metamorphism by virus/worm requires A high degree of metamorphism and similarity to other programs This is not trivial!

Conclusion  How practical is our detection method?  We “cheat” in several ways Use IDA to disassemble Viruses not embedded in other code Limited testing, small no. of files, etc.  But results appear to be robust If so, we can be “sloppy” (i.e., more efficient) and still get good results

The Bottom Line  Metamorphism for “good” For example, buffer overflow mitigation A little metamorphism does a lot of good  Metamorphism for “evil” For example, try to evade virus/worm signature detection Requires high degree of metamorphism and similarity to normal programs Not impossible, but not easy…

The Bottom Bottom Line  For metamorphic software, perhaps the inherent advantage lies with the good guys rather than the bad guys  All-too-often in information security, the advantage lies with the bad guys

References  X. Gao, Metamorphic software for buffer overflow mitigation, MS thesis, Dept. of CS, SJSU, 2005  P. Szor, The Art of Computer Virus Research and Defense, Addison-Wesley, 2005  M. Stamp, Information Security: Principles and Practice, Wiley InterScience, 2005  W. Wong, Analysis and detection of metamorphic computer viruses, MS thesis, Dept. of CS, SJSU, 2006  W. Wong and M. Stamp, Hunting for metamorphic engines, to appear in Journal in Computer Virology

Appendix Extra Materials

HMMs – Run Time of Training Process  5 to 38 minutes, depending on number of states n.

HMMs – Run Time of Classifying Process  to 0.4 milliseconds, depending on N and number of opcodes T.

AVG Anti-Virus Scanning Result