Network Structure Grant Schoenebeck, Aaron Snook, Fang-Yi Yu Sybil - - PowerPoint PPT Presentation

▶
network structure
SMART_READER_LITE
LIVE PREVIEW

Network Structure Grant Schoenebeck, Aaron Snook, Fang-Yi Yu Sybil - - PowerPoint PPT Presentation

May 02, 2023 585 likes •884 views

Sybil Detection Using Latent Network Structure Grant Schoenebeck, Aaron Snook, Fang-Yi Yu Sybil Attack An attack to compromise a recommendation systems by forging identities. Recommendation System How is that restaurant? Bad Good Good

slide-1
SLIDE 1

Sybil Detection Using Latent Network Structure

Grant Schoenebeck, Aaron Snook, Fang-Yi Yu

slide-2
SLIDE 2

Sybil Attack

  • An attack to compromise a

recommendation systems by forging identities.

slide-3
SLIDE 3

Recommendation System

How is that restaurant? Good Bad Good Good Good Bad Bad

slide-4
SLIDE 4

Sybil Can Manipulate the Opinion

Good Bad Good Good Good Bad Bad How is that restaurant?

slide-5
SLIDE 5

Activities and Profile Characteristics

  • Pros

– Proliferating signals to exploit – Practical benefits

  • Cons

– Cat and mouse game

slide-6
SLIDE 6

Structure of the Social Network

  • Pros

– Expensive signal to forge

  • Cons

– Stringent conditions

slide-7
SLIDE 7

Assumptions on Network Topology

  • Assuming distinct ability

– Honest nodes: Well-mixed networks – Sybil: Limited connection to the honest

  • Empirical results [Alvisi 2013]

– Social networks don’t have fast mixing time – Sybil are accepted as friends much higher than anticipated

slide-8
SLIDE 8

Alternative Assumptions

Previous Assumptions

  • Honest nodes:

– Well-mixed networks

  • Sybil:

– Limited connection to the honest

Goal

  • Recover all honest agents

Our Assumptions

  • Honest nodes:

– ‘locally’ dense in low dimensional space

  • Sybil:

– relax to constant fraction of honest agent would be compromisable

Goal

  • core space: a whitelist of nodes
slide-9
SLIDE 9

Alternative Assumptions

Previous Assumptions

  • Honest nodes:

– Well-mixed networks

  • Sybil:

– Limited connection to the honest

Goal

  • Recover all honest agents

Our Assumptions

  • Honest nodes:

– ‘locally’ dense in low dimensional space

  • Sybil:

– relax to constant fraction of honest agent would be compromisable

Goal

  • core space: a whitelist of nodes
slide-10
SLIDE 10

Low Dimensional Latent Metric Space

  • Intuition

– Metrics space encodes the similarity between agents

  • Well-regarded network models

– Watts-Strogatz model: ring – Kleinberg’s small world model: lattices – Low distortion multiplex social network [Abraham2013]

slide-11
SLIDE 11

Our Low Dimensional Assumptions

  • Dimensionality

– Graph with pairwise distance – Requiring low doubling dimension

  • Density

– Every local region contains a random graph – Only require of constant fraction of nodes

  • How realistic are our assumptions

having ℝ𝑒as special cases

slide-12
SLIDE 12

Experiment Setups

  • Dataset Description

– Facebook – Twitter – Wiki-vote – Epinion

  • Implementation

– Use Spectrum embedding – Compute the core space

slide-13
SLIDE 13

Core Space in Facebook

  • Graph properties

– 4,039 nodes, 88,234 edges – Average degree 21.8

  • Core space

– Density > 10 – Connect to 𝑞 fraction of nearby nodes

slide-14
SLIDE 14

Core Space in Twitter

  • Graph properties

– 81,306 nodes, 1,768,149 edges – Average degree 21,75

  • Core space

– Density > 10 – Connect to 𝑞 fraction of nearby nodes

slide-15
SLIDE 15

Alternative Assumptions

Previous Assumptions

  • Honest nodes:

– Well-mixed networks

  • Sybil:

– Limited connection to the honest

Goal

  • Recover all honest agents

Our Assumptions

  • Honest nodes:

– ‘locally’ dense in low dimensional space

  • Sybil:

– relax to constant fraction of honest agent would be compromisable

Goal

  • core space: a whitelist of nodes
slide-16
SLIDE 16

Compromisable Agents

  • Idea

– Someone might accept all the friend requests

  • Honest nodes

– Most of the nodes are trustworthy – A random portion of nodes are compromisable

  • Sybils

– Cannot connect to trustworthy nodes

slide-17
SLIDE 17

Assumptions Summary

Assumptions Social network Sybils Previous Works Well-mixed Bounded connection to honest nodes Our Work Locally dense in low- dimensional space Only connection to compromisable nodes

slide-18
SLIDE 18

Detection Game

  • Original Graph
slide-19
SLIDE 19

Detection Game

  • Reveal the trustworthy and compromisable nodes
slide-20
SLIDE 20

Detection Game

  • Adversary try to add Sybil nodes into the networks
slide-21
SLIDE 21

Detection Game

  • Adversary try to add Sybil nodes into the networks
slide-22
SLIDE 22

Detection Game

  • Detection algorithm return a whitelist
slide-23
SLIDE 23

Detection Game

  • Detection algorithm return a whitelist
slide-24
SLIDE 24

Theorem

  • If the total number of Sybil nodes and Compromisable nodes is

smaller than some constant fraction the honest nodes, and the graph can be imbedded into locally dense low dimensional space, in the detection game for any adversary the detection algorithm can return a large whitelist without any Sybil

slide-25
SLIDE 25

A Toy Model

  • Network of honest nodes

– 1 dimensional unit circle – 𝑜 nodes uniformly placed – Well-connected within distance

1 log 𝑜

  • Limitation of Sybils

– Connects to Sybil or compromisable node – #Sybil = O(𝑜), #the Compromisable = O(𝑜)

slide-26
SLIDE 26

A Toy Model

  • Network of honest nodes

– 1 dimensional unit circle – 𝑜 nodes uniformly placed – Well-connected within distance

1 log 𝑜

  • Limitation of Sybils

– Connects to Sybil or compromisable node – #Sybil = O(𝑜), #the Compromisable = O(𝑜)

slide-27
SLIDE 27

What can Sybil do?

Connect to the compromisable Form its own network

slide-28
SLIDE 28

What should detection algorithm do?

Remove non-local edges Remove low degree nodes

slide-29
SLIDE 29

Future Work

  • Can we do better if we have information of compromisable

nodes?