| dc.description.abstract | Online auctions have become one of the most popular and convenient buying and selling
media in e-commerce. However, the amount of auction fraud increases with the popu-
larity of online auctions. This thesis examines one of the most severe types of auction
fraud, referred to as shill bidding, where fake bids are used to arti cially in
ate an item's
nal price. Shill bidding is strictly prohibited in online auctions because it forces honest
bidders to pay more for their products. Researchers have proposed several mechanisms
to detect shill bidding once an auction has nished. However, if shill bidding is not
detected during an auction, an innocent bidder (i.e., the winner of the auction) can
potentially be cheated by the end of the auction. Therefore, it is necessary to detect
and verify potential shill bidding in real-time (i.e., while an auction is in progress).
This thesis proposes and implements several novel techniques for combating shill
bidding in real-time. The e ectiveness of these techniques is shown through applica-
tions on simulated and commercial auction datasets. The research results include the
following:
1. Existing literature conducted on shill bidding detection and prevention has been
studied. The research gaps that remain as research challenges in this particular
domain have been identi ed. The research ideas on the detection and prevention
of shill bidding have been summarised from past literature. This will help future
researchers to develop more trustworthy online auction environments.
2. Testing the e ectiveness of newly proposed fraud countermeasures is a very chal-
lenging task. The reason is the lack of auction data availability as commercial
online auction houses (e.g., eBay, Yahoo! Auctions, etc.) do not share their data
for privacy concerns. Furthermore, there is no way to evaluate the performance of
fraud detection mechanisms. To accomplish these research goals, an online auc-
tion system, referred to as uAuction, has been developed. uAuction has been
used to test the e ectiveness of newly proposed shill bidding detection algorithms.
3. A real-time shill bidding detection algorithm, referred to as the Live Shill
Score (LSS), has been introduced to identify shill bidding while an auction is in
progress. Experimental analysis shows that the LSS algorithm can detect potential
shill bidders at the 80% duration mark of an online auction in most cases, which
is 20% quicker at detecting potential shill bidding than the existing published
proposals.
4. As real-world auction data contains a large number of auctions and participants,
the LSS algorithm has been extended to multiple auctions. The extended ver-
sion of real-time shill bidding detection algorithm has been referred to as
the multi-auction LSS. The algorithm has been implemented with a single seller.
Experimentation has been undertaken to determine how e ective the algorithm
is in identifying shill bidders across multiple auctions. However, some innocent
bidders were incorrectly classi ed as shill bidders when the algorithm considered
multiple auctions hosted by only a single seller. To avoid this misclassi cation,
the multi-auction LSS algorithm has been considered for multiple auctions con-
ducted by multiple sellers. Experimental results show that the algorithm is able
to accurately detect potential shill bidding across multiple auctions. Comparative
analysis shows that the multi-auction LSS with multiple sellers performs better
than the multi-auction LSS with a single seller and the LSS algorithm.
5. Detecting potential shill bidders becomes more di cult when a seller involves two
or more bidders and forms a collusive group to commit shill bidding in his/her own
auction(s). The reason is colluding shill bidders can distribute the work evenly
among each other to collectively reduce their chances of being detected. An o ine
approach to detect colluding shill bidders in online auctions based on machine
learning techniques has been proposed. The solution is referred to as the Collusive
Shill Bidding Detection (CSBD). The CSBD algorithm has been implemented and
applied on both simulated and commercial auction datasets. Experimental results
show that the CSBD algorithm can detect potential colluding shill bidders. Later,
the algorithm has been extended to a real-time approach, referred to as the real-
time CSBD algorithm. The real-time CSBD has been implemented and applied
on simulated and commercial auction datasets. Experimental analysis shows that
the algorithm is able to highlight potential colluding shill bidders in real-time.
6. Shill bidding detection becomes more complicated when a seller registers multiple
fake accounts and engages shill bidders in the auctions operated by those fake seller
accounts. Such behaviour is referred to as multiple seller collusive shill bidding,
where a seller can distribute the risk between the colluding sellers and reduce their
chance of identifying any particular seller (i.e., a shill bidder). There is no lit-
erature available on the detection of multiple seller collusive shill bidding
in real-time. A basic real-time detection algorithm using social network analy-
sis (SNA) has been proposed for identifying multiple seller collusive shill bidding
while an auction is in progress. The algorithm has been referred to as the Multi-
ple Seller Collusive Shill Bidding Detection (MSCSBD). The algorithm has been implemented and applied on synthetically generated auction dataset. Experimen-
tal results show that the algorithm behaves reasonably and can identify potential
colluding sellers who engage in shill bidding while an auction is in progress. | |
