CISC 879 - Machine Learning for Solving Systems Problems
- 1. Presented by: Sandeep Dept of Computer & Information Sciences University of Delaware Detection of unknown computer worms based on behavioral classification of the host Robert Moskovitch ,Yuval Elovici ,Lior Rokach
- 2. Worms Worms are considered malicious in nature Worms propagate actively over a network, while other types of malicious codes, such as viruses, commonly require human activity to propagate Viruses infect a file (its host), a worm does not require a host file .
- 3. What do Antivirus Packages do ? Antivirus software packages inspect each file that enters the system, looking for known signatures which uniquely identify an instance of known malcode Polymorphism and metamorphism are two common obfuscation techniques used by malware writers Polymorphic virus obfuscates its decryption loop using several transformations, such as nop- insertion, code transposition
- 4. Obfuscation Techniques Metamorphic viruses attempt to evade detection by obfuscating the entire virus. When they replicate, these viruses change their code in a variety of ways, such as code transposition, substitution of equivalent instruction sequences, change of conditional jumps, and register reassignment
- 5. Example Virus Code : Morphed Virus Code(From Chernobyl CIH1.4) Loop : Loop : pop ecx Loop: pop ecx pop ecx nop nop jecxz SFModMark jecxz SFModMark jmp L1 mov esi , ecx xor ebx , ebx L3: call edi mov eax , 0d601h beqz N1 xor ebx , ebx Pop edx N1: mov esi , ecx beqz N2 Pop ecx nop N2: jmp Loop Call edi mov eax ,0d601h jmp l4 pop edx L2: nop pop ecx mov eax , 0d601h nop pop edx Xor ebx , ebx call edi pop ecx beqz N1 Xor ebx , ebx nop N1: mov esi , ecx beqz N2 jmp L3 jmp L2 N2: JMP loop L1: jecxz SFModMark L4:
- 6. Current Methods Existing methods rely on the analysis of the binary for the detection of unknown malcode. Some less typical worms are left undetectable. Therefore an additional detection layer at runtime is required
- 7. Proposed Approach Malicious actions are reflected in the general behavior of the host. By monitoring the host, one can inexplicitly identify malcodes. A classifier is trained with computer measurements from infected and not infected computers.
- 8. Contributions of the Paper Machine learning techniques are capable of detecting and classifying worms Using feature selection techniques to show that a relatively small set of features are sufficient for solving the problem without sacrifice accuracy. Empirical results from an extensive study of various machine configurations suggesting that the proposed methods achieve high detection rates on previously unseen worms.
- 9. Train and Test Phase
- 10. Dataset creation Lab network consisted of seven computers, which contained heterogenic hardware, and a server computer simulating the internet. Used the windows performance counters and Vtrace which enable monitoring system features A vector of 323 features for every second. Choose worms that differ in their behavior, from among the available worms
- 12. Feature selection methods Chi-Square Gain Ratio Relief Features’ ensemble : fi is a feature, filter is one of the k filtering (feature selection) methods.
- 13. Consolidating features from different environments:Averaged and Unified
- 14. Feature Sets
- 15. Classification algorithms Decision Trees, Naıve Bayes, Bayesian Networks Artificial Neural Networks
- 17. Experiment I Each classifier is trained on a single dataset i and tested on each one ( j ) of the eight datasets. Eight corresponding evaluations were done on each one of the datasets, resulting in 64 evaluation runs. When i = j , 10 fold cross validation, in which the dataset is randomly partitioned into ten partitions and repeatedly the classifier is trained on nine partitions and tested on the tenth.
- 18. Experiment I (Contd) Each evaluation run (out of the 64) was repeated for each one of the combinations of feature selection method, classification algorithm, and number of top features. Each evaluation run was repeated for the 33 feature set described earlier 132 (four classification algorithms applied to 33 feature sets) evaluations (each comprises 64 runs), summing up to 8448 evaluation runs.
- 19. Results
- 20. Results(Contd)
- 21. Results(Contd)
- 22. Experiment II Classifiers based on part of the (five) worms and the none activity, and tested on the excluded worms (from the training set) and the none activity Training set consisted of 5 − k worms and the testing set contained the k excluded worms, while the none activity appeared in both datasets. This process repeated for all the possible combinations of the k worms (k = 1–4). The Top20 features, which outperformed in e1 were used
- 23. Results
- 24. Conclusion Q1: In the detection of known malicious code, based on a computer’s measurements, using machine learning techniques, what is the achievable level of accuracy? Q2: Is it possible to reduce the number of features to below 30, while maintaining a high level of accuracy
- 25. Conclusions(Contd) Q3: Will the computer configuration and the computer background activity, from which the training sets were taken, have a significant influence on the detection accuracy? Q4: Is the detection of unknown worms possible, based on a training set of known worms?