Detecting Fraudulent Personalities in Networks of Online Auctioneers Duen Horng (“Polo”) Chau...
-
date post
20-Dec-2015 -
Category
Documents
-
view
216 -
download
1
Transcript of Detecting Fraudulent Personalities in Networks of Online Auctioneers Duen Horng (“Polo”) Chau...
Detecting Fraudulent Personalities in Networks of Online Auctioneers
Duen Horng (“Polo”) ChauShashank Pandit
Christos Faloutsos
School of Computer ScienceCarnegie Mellon
PKDD ’06, Berlin, Germany
“Detecting Fraudulent Personalities in Networks of Online Auctioneers” Polo Chau, Shashank Pandit, Christos Faloutsos
School of Computer Science, Carnegie Mellon University, USA2
Duen Horng (Polo) CHAU(author of these foils –used with his permission)
Shashank PANDIT
Online auctions:very popular
“Detecting Fraudulent Personalities in Networks of Online Auctioneers” Polo Chau, Shashank Pandit, Christos Faloutsos
School of Computer Science, Carnegie Mellon University, USA4
Why care about auction fraud?
REASON 1: it’s a serious problem 14,500 complaints received by Internet Crime
Complaint Center in USA in 2005 Average loss per incident: > US$385
REASON 2: it’s a hard problem No systematic approaches, until now.
“Detecting Fraudulent Personalities in Networks of Online Auctioneers” Polo Chau, Shashank Pandit, Christos Faloutsos
School of Computer Science, Carnegie Mellon University, USA6
$$$
Example of an online auction
Seller Buyer
A TransactionWhat if something goes BAD in the transaction?
Non-delivery fraud
Very Common
We focus on dealing with it.
“Detecting Fraudulent Personalities in Networks of Online Auctioneers” Polo Chau, Shashank Pandit, Christos Faloutsos
School of Computer Science, Carnegie Mellon University, USA7
Buyer
Feedback score: 15
$$$
Feedback on an online auction
Seller
Feedback score: 70
A Transaction
+ 1 = 71 - 1 = 14
Each user has a feedback score (= # positive feedback - # negative feedback)
How to game the feedback system?(and how to guard against gaming?)
“Detecting Fraudulent Personalities in Networks of Online Auctioneers” Polo Chau, Shashank Pandit, Christos Faloutsos
School of Computer Science, Carnegie Mellon University, USA10
Do fraudsters follow some patterns when they boost reputation?
Too “wasteful”; whole (near) clique will be lost
Will never deliverWill never deliver
Will never deliver
“Detecting Fraudulent Personalities in Networks of Online Auctioneers” Polo Chau, Shashank Pandit, Christos Faloutsos
School of Computer Science, Carnegie Mellon University, USA11
They form near-bipartite cores
The bad guys (humans) create 2 types of users
Accomplice Trade mostly with
honest users Looks legitimate
Fraudster Trade mostly with
accomplices Don’t trade with other
fraudsters
“Detecting Fraudulent Personalities in Networks of Online Auctioneers” Polo Chau, Shashank Pandit, Christos Faloutsos
School of Computer Science, Carnegie Mellon University, USA12
Why near-bipartite cores?
Allow accomplices to be reused
Hard to discover because they look very legitimate
Fraudsters will get voided, but only one at a time
Will never deliver
Research Goal: Detect the suspicious near-bipartite cores
Our Approach:
Use the belief propagation (BP) algorithm
“Detecting Fraudulent Personalities in Networks of Online Auctioneers” Polo Chau, Shashank Pandit, Christos Faloutsos
School of Computer Science, Carnegie Mellon University, USA14
Belief Propagation (BP) algorithm
Efficient way to solve inference problems based on passing local messages E.g. Used in early vision problem, such as image
restoration
Useful for our problem as well! (Thanks to John Lafferty for pointers!)
Details
“Detecting Fraudulent Personalities in Networks of Online Auctioneers” Polo Chau, Shashank Pandit, Christos Faloutsos
School of Computer Science, Carnegie Mellon University, USA15
Belief at each node
Probability being fraudster
Probability being accomplice
Probability being honest
Details
“Detecting Fraudulent Personalities in Networks of Online Auctioneers” Polo Chau, Shashank Pandit, Christos Faloutsos
School of Computer Science, Carnegie Mellon University, USA16
Example Message passing is iterative.Beliefs keep being updated, until equilibrium is reached
A
C
B
E
D
Details
“Detecting Fraudulent Personalities in Networks of Online Auctioneers” Polo Chau, Shashank Pandit, Christos Faloutsos
School of Computer Science, Carnegie Mellon University, USA17
Edge Compatibility Function
The function specifies how the belief of a node affects its neighbors (in our case, it captures the bipartite core structure)
In our context, the function can be represented as the following matrix:
Entry(i, j) = probability that a node is in state j given that it has a neighbor in state i
Details
“Detecting Fraudulent Personalities in Networks of Online Auctioneers” Polo Chau, Shashank Pandit, Christos Faloutsos
School of Computer Science, Carnegie Mellon University, USA18
Belief propagation -- mathematically Details
Message to send out from a node based on its belief
Belief at a node
Edge compatibility function
Experiments
“Detecting Fraudulent Personalities in Networks of Online Auctioneers” Polo Chau, Shashank Pandit, Christos Faloutsos
School of Computer Science, Carnegie Mellon University, USA20
FraudstersAccomplices Honest
ConfirmedFraudsters
Effectiveness on real data Real data from eBay
60K users; 1M edges (more data – 12Gb/day…)
“Detecting Fraudulent Personalities in Networks of Online Auctioneers” Polo Chau, Shashank Pandit, Christos Faloutsos
School of Computer Science, Carnegie Mellon University, USA21
“Detecting Fraudulent Personalities in Networks of Online Auctioneers” Polo Chau, Shashank Pandit, Christos Faloutsos
School of Computer Science, Carnegie Mellon University, USA22
In the news
(thanks to Byron Spice) WSJ online AP LA Times San Jose Mercury News KDKA USA Today
“Detecting Fraudulent Personalities in Networks of Online Auctioneers” Polo Chau, Shashank Pandit, Christos Faloutsos
School of Computer Science, Carnegie Mellon University, USA23
Industrial etc interest
e-bay Symantec (thanks to Bill Courtright of PDL) ‘Belgian police’ -> probably fraudster in
disguise (!?)
“Detecting Fraudulent Personalities in Networks of Online Auctioneers” Polo Chau, Shashank Pandit, Christos Faloutsos
School of Computer Science, Carnegie Mellon University, USA24
Conclusions
Method to detect auction fraud Use belief propagation Detect the near bipartite cores
Evaluated with real eBay data and synthetic data