1. Towards Network Containment in Malware Analysis Systems Authors: Mariano Graziano, Corrado Leita, Davide Balzarotti Source: Annual Computer Security Applications Conference (2012) Reporter: MinHao Wu

2. Outline Introduction Malware analysis and containment Protocol inference System overview Evaluation Conclusion

3. Introduction Dynamic analysis is a useful instrument for the characterization of the behavior of malware. The most popular approach to perform dynamic analysis consists in the deployment of sandboxes The result of the execution of a malware sample in a sandbox is highly dependent on the sample interaction with other Internet hosts. The network traffic generated by a malware sample also raises obvious concerns with respect to the containment of the malicious activity.

4. Malware analysis and containment

5. Protocol inference

6. System overview Traffic Collection ◦ By running the sample in a sandbox or by using past analyses Endpoint Analysis ◦ Cleaning and normalization process Traffic Modeling ◦ Model generation (two ways: incremental learning or offline) Traffic Containment ◦ Two modes (Full or partial containment)

7. Traffic Collection running a network sniffer while the sample is running in the sandbox. several online systems allow users to download in our experiments we limited the malware analysis and the network collection time to five minutes per sample.

8. Endpoint Analysis cleaning and normalizing the collected traffic to remove spurious traces and improve the effectiveness of the protocol learning phase the cleaning phase mainly consists in grouping together traces that exhibit a comparable network behavior

9. Traffic Modeling

10. Containment Phase

12. EVALUATION All the experiments were performed on an ◦ Ubuntu 10.10 machine running ScriptGen, Mozzie, and iptables v1.4.4. ◦ To perform the live experiments, we ran all samples in a Cuckoo Sandbox [6] running a Windows XP SP3 virtual machine.

13. Results of the Offline learning Experiments Fast flus

14. Results of the Incremental learning Experiments

15. Tested samples: ◦ 2 IRC botnets, 1 HTTP botnet, 4 droppers, 1 ransomware, 1 backdoor and 1 keylogger Required network traces ranging from 4 to 25 (AVG 14) DNS lower bound (6 traces)

16. CONCLUSIONS The benefits of the large-scale application of similar techniques are significant ◦ old malware samples ◦ in-depth analyses of samples