Sharing Files on Peer-to-Peer Networks based on Social Network

Fabrice Le Fessant

INRIA  Saclay-Île de France 23 Février 2009

1 / 14 Social networks are less vulnerable to attacks by spies

⇒ peer-to-peer network + social network ⇒ less vulnerable to censorship... maybe

Fighting Censorship

Standard peer-to-peer networks are vulnerable → Spies can locate content providers

States and Big Companies already have spies on P2P networks

2 / 14 Fighting Censorship

Standard peer-to-peer networks are vulnerable → Spies can locate content providers

States and Big Companies already have spies on P2P networks

Social networks are less vulnerable to attacks by spies

⇒ peer-to-peer network + social network ⇒ less vulnerable to censorship... maybe

3 / 14 Peer-to-Peer + Social Network

Each user allows his computer to connect only to his friends' computers

What can we do with such a network ?

4 / 14 My main requirement: • Content providers should remain unknown • while Links (friends) are not hidden • ISP must share such information with spies

File-Sharing

Three main operations: Resource discovery: nd which resources exist Resource localization: nd which resources are available and where Resource access: download the resources

5 / 14 File-Sharing

Three main operations: Resource discovery: nd which resources exist Resource localization: nd which resources are available and where Resource access: download the resources

My main requirement: • Content providers should remain unknown • while Links (friends) are not hidden • ISP must share such information with spies

6 / 14 Related Work

Restrict unstructured networks to social based networks Gnutella Simple ooding (Turtle) FreeNet Local Routing with replication GnuNet Flooding with probabilistic shortcuts

7 / 14 The Turtle Network

• Gnutella on a Social Network • Search by ooding the neighbours → Tree Structure → Reply from leaves to root

• Too expensive in bandwidth → But interesting simulations on Orkut

8 / 14 Popularity of Users

in Orkut trace: 106,481 users in 2004 9 / 14 Connectivity

Users linked to the main component

10 / 14 Size of largest Component

11 / 14 Distribution of Path Lengths

12 / 14 Orkut Crawl (2004)

Users by Popularity 100 Orkut 2004 10 100

10

1 Number of neighbours

0.1 0 20000 40000 60000 80000 100000 120000 CDF of users

13 / 14 Orkut Crawl (2007)

Users by Popularity 100000 Orkut 2007/10/01 10 10000 100 500

1000

100

10 Number of neighbours 1

0.1 0m 0.5m 1m 1.5m 2m 2.5m 3m 3.5m CDF of users

14 / 14 Orkut Crawl (2009)

Users by Popularity 100000 Orkut 2009/02/04 10 10000 100 500 32000 1000

100

10 Number of neighbours 1

0.1 0m 0.5m 1m 1.5m 2m 2.5m 3m 3.5m 4m 4.5m 5m CDF of users

15 / 14 Orkut Crawl (2009)

Crawler Performance 5m 100k 4.5m 90k 4m 80k 3.5m 70k 3m 60k 2.5m 50k 2m 40k 1.5m 30k Number of Queued Users 1m 20k Number of Crawled Users

0.5m Queued 10k Crawled 0m 0k 0 50 100 150 200 250 Hours

16 / 14 Reliability of Orkut Crawls

Real topology of a social network ? • First crawl (2004): crawled users and queued users probably mixed • Second crawl (2007): only links between crawled users were kept • Third crawl (2009): uses advertized popularity for queued users

17 / 14 FreeNet

Anti-censorship network for small documents • Can be limited to friends • Documents are encrypted • Routing: • Depth First Search • Closest Neighbour rst • Access: • Document Replicated on way back

18 / 14 GnuNet

Same goal as FreeNet, but: • Routing: • Limited-Breadth First Search • Shortcut system (not with friends !) • Credit System (anti-ooding) • Optimizations are weaknesses

19 / 14 Can we do better ?

• Designed for small les • FreeNet and GnuNet: • Expect stronger attackers: • Trac analysis • Moving connectivity • Turtle: • Unstructured

20 / 14 Queries or Publications

• Current Systems focus on whole-system queries • My position: • Try to structure the network • Half query/half publication (like DHT...) • Whole-system publications • File-type optimizations/issues • New problems: • Real network must be a sub-network of social one • Use source-coding for ecient big downloads ⇒ Requires replica localization

21 / 14