Shilling is a kind of fraudulence colluded by shilling sellers and bidder. The seller posts an auction; the fake bidders place the bid but never win. Finally, the legitimate buyer will win the bid with an inflated price. Currently, most online auctioning marketplaces, such as eBay, use feedback based reputation system to guarantee the trustworthy on the pseudonymous environment. But this mechanism can not alarm the buyers about shilling because the feedback from previous buyers mostly focuses on quality of the commodity and whether the seller delivers the products promptly. It does not provide information about the final price and no analysis on the relationship between a seller and other bidders. This paper applies three statistical models to measure the behavior of sellers and bidders, design a reputation score schema to identify suspicious fraud. At last, they demonstrated it on eBay and identify potentially dishonest sellers. The paper designs model N to identify the sellers with abnormal high number of bids. In a fixed period of time, the demand for a particular category commodity is limited. No matter how many auctions the seller posts, the number of buyers is almost fixed. What's more, if a seller posts many auctions of similar commodities, it also distracts the buyer’s focus. But for a shill seller, because there are shill bidders participating in most of its auctions no matter how many auctions it posts, the shill sellers’ average number of bids often abnormally high. But this model would introduce false positive. If the seller often posts competitive auction, the bid number might be often high. To rectify the false positive of Model N, the author designs Model M, which analyzes the relationship between the average bid number and the minimum starting bid. Intuitively, the lower starting bid should attract more bids, otherwise fewer bids. But for a shill seller, even its minimum starting bid is high, because a number of fake bids always participates its auction, its bid number will be still very high. So its goal is to identify sellers with abnormally high number of bids and high minimum starting bid. The author defines RMB: relative minimum starting bid, which is related to the ratio between the minimum starting bid and the winning bid. The larger RMB means the lower starting bid. They plot the diagram of average RMB of a seller and the average bid number in seller’s auction and find the intuitive inverse correlation, that is, the lower minimum starting results in more bids. But false positive still exists, for a competitive auction with a very low minimum starting bid, the bid number might exceed the up bound. If the commodity is very good, it still can attract many bids even with a high starting bid. To further rectify the model, the author designs Model P, which analyzes the bidder's profile of a seller. They observe that fake bidders always participate in a particular seller's auction many times but almost never win. Model P is intended to identify such bidders. Model P is based on an assumption: that legitimate active bidders win auctions at a rate similar to less active bidders. That means, the win number should be proportional to the participant number. The more auctions you attend, the more auctions you should win. For each seller, the author plots a bidder presence curve and bidder win curve. The abnormal sellers are those sellers whose bidder presence and win curve has a big gap. That means it has a group of bidders that repeatedly participate and lose. The author also notices that a single model would introduce much false positive and can not handle special cases, so they design a reputation score to combine the three models to detect potential shilling sellers more efficiently. The paper’s reputation system is a good complement to existing reputation system for online auctioning by alarming buyers of suspicious shilling. Because it is based on seller and buyers’ business behavior, rather than buyers’ feedback, even if the seller’s feedback reputation is very excellent, it still works. This reputation system is also applicable because it only needs simple data, such as auction/bids number. Although there are other more sophisticated methods to detect shill, those methods need complicated data which makes the data analysis difficult. Although these three models can identify potential shill, it is not hard to beat them. First, the fraudulent seller can register a number of bidders and distribute them among different auctions, reduce bidders’ presence. Thus, the bidder presence curve and win curve will match to a particular extend. Next, the seller can decrease minimum starting bid and use few shill bids per auction. These two steps can beat model M and N substantially. As long as two models are beaten, the reputation score can not identify the potential shill. What’s more, the reputation score is not plausible. Each model has the same weight, but actually Model P “catches” the key feature of the shilling seller and bidders. It should have more weight than the other two. A more reasonable weight solution should be designed. Besides revise the reputation score, I think more measurements could improve the reputation system's performance. There are three more observations. First, shill bidder prefer to bid earlier because late bidding increase the risk to win the bid while legitimate bidders have no enough time to respond fake bid. We can propose a statistical model on bid time. Second, as we said before, the shill bids have an inflated price, which is comparable to market price. So, I think we can calculate the ratio of the winning bid and market price just when the buyer won the bid, because the market price will change as time goes on. The shill seller’s ratio would be close to 1. Third, shill sellers and the bidders have strong associations in other categories if the seller puts auctions in various categories. Legitimate buyers usually buy a particular category commodity offered by a seller. If the seller also sells other commodities, the legitimate buyer is seldom interested in at the same time. But the fake bidders will follow the seller and place bids in different categories. We can measure this association. But this method is a little complicated in data collection and less effective because some high volume sellers usually concentrate on only one category commodity. The reading group's vote is following: Reject: 4 Weak Reject: 2 Weak Accept: 3 Accept: 0 Most of us think it is a good paper. But it is not a typical security paper. It is not suitable for CCS. Actually, 2/3 reference of this paper is from business and management.