![](https://csdnimg.cn/release/download_crawler_static/3466822/bg4.jpg)
Key insight: Secure routing protocols only deal with one half of
the problem. A clever export policy can attract as much traffic as a
bogus announcement.
Q: How robust are the statements you make?
A: We ran the experiments twice on two different data sets, includ-
ing UCLA’s Cyclops data sets. Trends from UCLA and CAIDA
datasets are the same.
Q: Have you compared the effectiveness of the various models
versus their deployability? Now that you know which ones are ef-
fective, and at least almost, which ones are deployable, where do
we go from here?
A: We completely ignored deployablity issues and implementabil-
ity issues and so on. Currently working on the deployability as-
pects.
Understanding Block-level Address Usage in the Visible
Internet
Xue Cai (USC/Information Sciences Institute), John Heide-
mann (USC/Information Sciences Institute)
What can simple observations about the Internet say?
Objective: Existing work provides little insight into the edge of
the Internet and the use of the IPV4 address space.
Methodology: Use active probing, pattern analysis, clustering
and classification.
Key insight: Provide information about how effectively network
address blocks appear to be used, provide new measurements about
dynamically managed address space, and show that low-bit rate last
hops are often underutilized.
Q: Have you done any work using IPV6 data? You seem to have
only IPV4?
A: No, we have IPV4 results only; probably in the future we will
look into V6.
Q: How reliable is it to utilize ping response for detection? Will it
work?
A: We didn’t do anything about it here.
Session 4: Privacy
Report by Fang Yu, Ohio State (yufa@cse.ohio-state.edu)
Privacy-Preserving P2P Data Sharing with OneSwarm
Tomas Isdal (University of Washington), Michael Piatek (Uni-
versity of Washington), Arvind Krishnamurthy (University of
Washington), Thomas Anderson (University of Washington)
Different types of data is being shared: public, private, public
but without attribution. This talk focuses on “public but without
attribution” data.
P2P network is good for content sharing. It is decentralized with
no central authority to change the rules or interfere with the trans-
fers. Their previous work has shown that current P2P systems are
very easy to monitor.
Why isn’t Tor good enough?
• Needs clients to be public
• Performance issue: high latency
Why isn’t Freenet good enough?
• Performance is even worse than Tor
OneSwarm:
• Had thousands of downloads
• Uses social network to establish friendship (P2P) relationship,
e.g., using Gmail account.
• Peers forward request and content over multiple hops, so you are
never sure if the neighboring node is the true content requester
or server.
Threat model: Attacker cannot observe traffic that she doesn’t
know. Attacker cannot inject traffic to arbitrary location in the net-
work.
Finding peers:
• Peers can be looked up in a DHT.
• Small number of friends makes it difficult to get service, get
good performance and peer can be vulnerable to attack. Solution
to this is to add community service so that P2P relationships can
be formed between untrusted nodes.
Resilience to timing attacks: This is dealt with by having same
behavior to the repeated request.
Search content:
• Search is flooded and delayed at each hop. Cancel message is
sent to cancel the search message once content is found.
• But there is no guaranteed every request is satisfied.
• Two types of searches: Hash (160 bits) search or content search
(plain text).
To deal with long paths, the solution is to use multi-paths. This
also provides resilience to failure on the path.
To transfer data, OneSwarm uses a modified version of BitTor-
rent. Also swarming download mode is used so that transfers for
the same content is grouped.
Timing attack: Attacker monitors the delay for search and re-
sponse. If the delay is small, the next-hop neighbor is suspected.
The results show that even with an extremely large number of at-
tackers, the attackers are still not able to narrow down to a single
sender.
Performance:
• 20MB file transferred between 120 PlanetLab nodes. OneSwarm
is much faster than Tor.
• Multipath has a huge impact on improving the transfer time.
Q: Is OneSwarm similar to Gnutella? Due to multi-hop, the per-
formance will be really bad?
A: Multi-hop is over a social overlay. Performance is worse.
Q: What is relied on for not knowing the source?
A: Through delay and multi-hops.
Q: Then this can be modeled using random variables? With longer
path, there can be leak of information?
A: It might be possible to estimate how far the source is.
Q: Is 20 MB too small? Why not try larger content?
A: Run this over Tor, not big of burden on the network. Same
size for different protocols. Have problem to get complete 20MB
transferred. 5 MBs then extrapolated.
Q: All nodes are stable?
A: Yes, all nodes are like a flash crowds.
Q: Did you experiment with node churn?
A: No.
Differentially-Private Network Trace Analysis
Frank McSherry (Microsoft Research), Ratul Mahajan (Mi-
crosoft Research)
Objective: Is it possible to do network analysis while achieving
“differential privacy” guarantee?
ACM SIGCOMM Computer Communication Review
Volume 41, Number 1, January 2011