Preserving Caller Anonymity in Voice-over-IP Networks

1. Preserving Caller Anonymity inVoice-over-IP Networks Mudhakar Srivatsa, Ling Liu and Arun Iyengar Presented by Mounica Atluri

2. Agenda • Voice-over-IP • Attacks • Proposed solution • Experimental Evaluation • Conclusion

3. Voice and data communication • Data transmission through Public switched telephone network • Uses Circuit switched networks • Expensive

4. What is VoIP ? • We see people talking through Skype, Vonage, instant messengers • Technology behind is called VoIP • Transmission of voice traffic over  IP-based networks • Sounds are recorded and compressed • Benefit of VoIP: Very economical

5. VoIP Requirements • Caller anonymity and QoS • Existing approaches use Mix networks • Mix networks route traffic through nodes with random delays and random routes • For example, Onion routing

6. VoIP Requirements • Other examples are Tor, Freedom and Tarzan • Mix networks cannot accommodate the QoS requirement • Low latency apps are vulnerable to timing attacks

7. Protocols in VoIP • Uses RTP for data transmission • Route Set Up protocol for call set up and termination

8. Route set up protocol • Operates in four steps • initSearch: initiates a route set up request • processSearch: processes a route set up request • processResult: processes the results of a route set up request • finSearch: concludes the route set up procedure

9. initSearch • src initiates a request by broadcasting

10. processSearch • If p receives a request from q, it checks if the sipurl is the url of the client connected to p. p

11. processResult • If p receives result (searchId, q), it searches for <searchId, sipurl, prev>, adds <sipurl, q> and forwards result to prev p

12. finSearch • If src receives result, it adds <dst, q> to its routing table q

13. Security features of Route setup protocol • Encryption with shared symmetric key • Exposes dst (through dst.sipurl) • dst adds a random delay • src or dst can be inferred if all of their neighboring nodes are malicious

14. Caller Identification attacks • Triangulation based timing attacks • 3 steps in triangulation based timing attacks • Candidate caller detection: malicious nodes deduce a list of potential callers • Candidate caller ranking: malicious nodes associate a score with every potential caller • Triangulation: Colluding malicious nodes combine their sets to obtain more accurate list of callers.

15. Three timing attacks • Deterministic triangulation attack • Statistical triangulation attack • Differential triangulation attack

16. Deterministic triangulation attack • 2 assumptions • Link latencies are deterministic • All nodes are synchronized • 2 properties of route setup protocol • Protocol establishes shortest route between the src and dst • Node p that receives route set up request originated from src can estimate dist(src, p)

17. Deterministic triangulation attack • Candidate caller detection • Compute S(p) for all s ∈ S(p),

18. Deterministic triangulation attack • Candidate caller ranking • Compute the score • Triangulation • Compute the final score

19. Deterministic triangulation attack

20. Statistical triangulation attack • Link latencies are independently distributed • Length of a path P is given by • In candidate caller detection, p computes a set of Pareto-optimal distances to all nodes v • A set of path lengths d1, d2.. dm is Pareto-optimal if for all other path lengths d,

21. Statistical triangulation attack • A node v is marked as a candidate caller if • If link latencies follow Gaussian, the path latencies follow Gaussian too • Score of v can be computed as • For other any other distribution, use Chebyshev’s inequality to compute

22. Statistical triangulation attack • In Triangulation step, the aggregate score for a candidate caller v is computed

23. Differential triangulation attack • Eliminates time stamp ts from the route set up request • Malicious nodes can estimate the difference • In candidate caller detection, malicious node p computes statistical shortest distances to every other node v as

24. Differential triangulation attack • Statistical distance distpq[v] is given by distp[v] – distq[v] • v is a candidate caller if • If the link latency distribution is Gaussian, the score of v is given by • Finally, the average score for v is computed

25. Topology Discovery • Network topology should be known for Timing attacks • Achieved by ping and pong messages y´ pong(y´,x) ping(x,all) x y pong(y, x)

26. Evaluation of the Threat models • Experimental set up • A synthetic network with 1024 nodes • Topology was constructed using NS-2 topology generator • Node-to-node round trip times varies from 24ms-150ms with a mean of 74ms

27. Deterministic Triangulation • Number of suspects varies with number of malicious nodes • Epsilon should not be too small or large

28. Statistical Triangulation • More effective than deterministic when there are uncertainties in link latencies

29. Differential Triangulation • Statistical attack performs better if the clocks are synchronized • Differential triangulation can achieve a top-10 probability of 0.78 with only 10 malicious nodes

30. Topology Discovery • With m=20 and ttl=2, about 75% of the topology is discovered

31. Countering timing attacks • Latency perturbation • each node adds random delay • Random Walk Search Algorithm • Resilient to timing attacks but generates suboptimal routes • Hybrid route set up • Trade off anonymity with QoS

32. Random Walk Search Algorithm • Sends a search request to a randomly chosen neighbor • Two key properties • Markovian property • Random walker does not traverse the shortest path between any two nodes

33. Hybrid route setup protocols • Controlled Random Walk • Combination of two protocols • γlimits the length of random walk • Starts with random walk search • Switches to broadcast search with probability 1-γ q

34. Hybrid route setup protocols • Multi-Agent Random Walk • Similar to random walk • Src sends ω random walkers (ω >1) • Route is established when the first random walker reaches dst • Higher ωresults in optimal route latency • Vulnerable to triangulation based timing attack if src sends out random walkers at time t=0

35. Experimental evaluation • Performed on 1024-node synthetic VoIP network topology using NS-2 • Algorithms implemented using Phex: an open source Java based implementation of peer-to-peer broadcast based route set up protocol

36. Performance • Characterized by cost of messaging • QoS guarantees • Routes with latency<250ms satisfy QoS requirements • Larger route set up latency does not affect the quality of voice conversation

37. Optimal parameter settings • Attack resilience • 99% optimal parameter settings

38. Topology discovery • Only fraction of topology has been discovered • Top-10 probability for marw was 42% less, crw was 33% less and broadcast was only 9% less • Random walk protocols are more sensitive to topology

39. Conclusion • VoIP in becoming popular due to its advantages in cost and convenience • It is a major concern to provide anonymity to the clients • Threat models targeting callers’ anonymity are efficient • Even if a small fraction of network is malicious, the caller can be inferred accurately • It is difficult to trade QoS with anonymity

40. Questions??