1 / 25

Detection of Covert Channels through VPNs – Final Presentation

Noam Segev , Israel Chernyak , Evgeny Reznikov. Detection of Covert Channels through VPNs – Final Presentation. Supervisor: Gabi Nakibly, Ph. D. . Covert Channels over VPNs . Mission Statement. Create a covert channel detector that would function in the described scenario.

mele
Download Presentation

Detection of Covert Channels through VPNs – Final Presentation

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Noam Segev, Israel Chernyak, EvgenyReznikov Detection of Covert Channels through VPNs – Final Presentation Supervisor: Gabi Nakibly, Ph. D.

  2. Covert Channels over VPNs

  3. Mission Statement • Create a covert channel detector that would function in the described scenario. • The detector’s operation: • A learning period of clear traffic • Two traffic samples: • A clean sample • Traffic containing a covert channel • Goal: Correct classification of both samples

  4. Methods Description • We used four detection methods in the creation of the detector: • BLOSUM • PSPM • Learning Algorithm • Entropy-based approach

  5. BLOSUM and PSPM • Taken from the field of bioinformatics • BLOSUM (BLOcks of Amino Acid SUbstitution Matrix) is a substitution matrix used for sequence alignment of proteins. • PSPM(Position Specific Probability Matrix) is a substitution matrix used for sequence alignment of proteins. • The algorithms constructs a substitution matrix of probabilities for each amino acid to be present in certain positions in the sequence.

  6. BLOSUM algorithm • We break down the learning communication to 10 groups of 10. (total length of 100) • We defined the probability of a value to be • We define the probability of a couple of values to be • When receiving a new packet we compare to a packet from the original communication using the formula • The values checked can be packet size or packet delay

  7. PSPM • We break down the learning communication to 10 groups of 10. (total length of 100) • We defined the probability of a value to be • The values checked can be packet size or packet delay

  8. Learning Algorithm • There exists a range of weaker algorithms for covert channel detection exist. • Each weak algorithm is either less accurate or only good for detecting a certain type of covert channel. • We utilized a learning algorithm in an attempt to boost and combine the effectiveness of several of the weaker algorithms.

  9. Learning Algorithm • We used the C4.5 learning algorithm to combine three of the weaker algorithms: • Regularity detection • Histogram of packet times/sizes • Epsilon similarity

  10. Learning algorithm metrics • Regularity: • Histogram:

  11. Epsilon Similarity • Stores and sorts the list of all inter-arrival times between packets. • Pi – inter-arrival time i in the sorted list. • Epsilon similarity: the percentage of |Pi - Pi+1|/Pi that are smaller than the epsilon.

  12. Learning Algorithm – cont. • During the semester we compiled a collection of traffic samples created by 3 of the covert channel programs designed by previous teams, as well as some samples of randomly generated traffic with normal distribution of sizes and inter-arrival times. • The learning algorithm was given a training set of the answers all the above methods gave for each packet in the aforementioned traffic.

  13. Entropy-based approach • Entropy measures the amount of disorder in a system. • A covert channel injects information into certain communication metrics, therefore increasing the amount of order over these metrics. • By measuring the amount of entropy of a given channel over the above metrics we can try to deduce the existence of a covert channel.

  14. Entropy calculation • We used the entropy calculation methods presented in Gianvecchio &Wang ’07:

  15. Entropy variables • Our method calculates the entropy of the following variables: • Packet delay • Packet sizes • Combined (size & time) • Bursts (k-packet averages on packet size & delay) • Peaks (maxima points of packet sizes and delays)

  16. First Challenge • The challenge consisted of 3 simulations. • Each included: • A learning phase on clean traffic. • A detection phase on clean traffic to weed out false positives. • And a detection phase on traffic contaminated by the covert channel.

  17. First Challenge – results • In this challenge our detector hasn’t generated any output – defined as a negative detection result – due to an error (which was found only later) which placed the output statements in an unreachable “if” statement.

  18. First Challenge - conclusions • Due to the aforementioned error suffered by our program, the results of the first challenge were inconclusive. • We decided that we’d attempt to investigate the sensitivity factor of our methods (since we saw neither false nor true positives).

  19. Learning Algorithm – analysis • We hoped that the algorithm would detect a pattern indicating which of the detection methods should be trusted in which case. • In case it was needed, we intended to boost the algorithm with providing more information about the covert channel – statistical information about packet distribution, as well as the numerical values computed by the aforementioned methods.

  20. Learning algorithm – analysis cont. • Unfortunately, it turned out that the covert channels we chose to work with were mainly detected by the histogram method, which didn’t leave much room for maneuver with the learning algorithm.

  21. First Challenge - improvements • Several issues were detected in our program: • The aforementioned error which prevented output from being displayed. • In the BLOSUM method, there were miscalculations in the algorithm. • The entropy calculations, albeit correct, suffered from inefficiency, which forced us to reduce several parameters, affecting the accuracy. • Sensitivity factors were tweaked throughout the program.

  22. Second Challenge • The second challenge consisted of one simulation, including, as in the first challenge: • A learning phase on clean traffic. • A detection phase on clean traffic to weed out false positives. • A detection phase on traffic contaminated by a covert channel.

  23. Second Challenge - results • Unfortunately, after weeding out false positives, no detection was made. • After some investigation, we discovered that the BLOSUM method has, in fact, detected the covert channel, but due to another error, failed to report it.

  24. Second Challenge - conclusions • Further refinement of the detection methods’ sensitivity thresholds is necessary. • In the learning algorithm method, the chosen methods proved to be insufficiently robust. Additionally, the lack of covert channel communication samples further undermined our efforts.

  25. Future work • It would be interesting to see how the learning algorithm fares given a large amount of traffic samples, as well as stronger methods such as the entropy and BLOSUM methods we have implemented during this project.

More Related