"... In PAGE 67: ... So, we found no correlation between the contents of received spam messages and the original music newsletter site. After analysing content of all spam messages, we classified them into ten different categories depending on what the spam offered (see Table1 for results). In this experiment, the categories that have generated the most spam were Free offers and Loans, credit card offers and financial services .... In PAGE 82: ... Although this was the case, these programs did not communicate with servers on the Internet during the experiment session. Table1 . Identified ad-/spyware programs Name Host Adware Spyware Download Internet BroadcastPC M x x x X KeenValue K x x XX Morpehus M X x BargainBuddy I, K x x x TopMoxie L, M x x x Cydoor I, K x x X Gator I, K X x X SaveNow B XX BonziBuddy L x x Web3000 I x x ShopAtHomeSelect I XXX WebHancer K x x BrilliantDigital K x XX MoneyMaker L, M XXX Claria I, K x X iMesh I x X WeatherCast B x X CasinoOnNet L x MyBar I, K, M x New.... In PAGE 83: ...73 In Table1 , a detailed list of the retrieved ad-/spyware components can be found. As can be seen, the ad-/spyware components were divided into Adware respectively Spyware based on their actions.... In PAGE 83: ... This means that any possible ad-/ spyware traffic distributed via BHOs is highly problematic to detect since it may very well be ordinary browser traffic. In Table1 , we also included two programs, New.Net and FavoriteMan, even though they were not classified as neither adware nor spyware.... In PAGE 84: ... In most cases the data was not readable, meaning that it was either encrypted or in a format not graspable. This is also an explanation to why we could confirm only two spyware programs (see Table1 ). Although most traffic data was not in clear text, we were able to extract and interpret some of the contents.... In PAGE 96: ...2.1 Experiment 1 A detailed list of the identified spyware programs is presented in Table1 . After having analysed the captured data, we concluded that all file-sharing tools con- tained spyware.... In PAGE 96: ... In addition to these findings, we also discovered that all file-sharing tools contained spyware that were involved in Internet communication. As can be seen in Table1 , the retrieved spyware components were divided into Adware and Spybot based on their operations. We also included a cate- gory called Download because some of the components allowed for further software and/or updates to be downloaded and installed.... In PAGE 97: ... Finally, spyware programs had an imperative effect on the amount of net- work traffic generated by the file-sharing tools. More specifically, there was a 48 Table1 . Identified spyware programs Name Host Adware Spybot Download Internet BroadcastPC M x x x X KeenValue K x x XX Morpehus M X x BargainBuddy I, K x x x TopMoxie L, M x x x Cydoor I, K x x X Gator I, K X x X SaveNow B XX BonziBuddy L x x Web3000 I x x ShopAtHomeSelect I XXX WebHancer K x x BrilliantDigital K x XX MoneyMaker L, M XXX Claria I, K x X iMesh I x X WeatherCast B x X CasinoOnNet L x MyBar I, K, M x New.... In PAGE 98: ...ith KaZaa. So, in contrast to KaZaa, installing a clean file-sharing tool (i.e., KaZaa Lite K++) caused marginal impact to system consumption and network bandwidth. However, due to the occurrence of spyware in file-sharing tools (see Table1 ), users with several such applications installed will, as a result of aggre- gate spyware activity, suffer from a continuos system and network degrading. 4 Discussion Based on the findings in Section 3, we can conclude that spyware programs exist, that they engage themselves in Internet communication, that they transmit user data, and that their existence have a negative impact on system and network capacity.... ..."

"... In PAGE 6: ... Features obtained by running other named-entity classifiers (with different tag sets): HMM, MaxEnt and RRM output on the 32-category, 49-category and MUC data sets.9 Table3 presents the mention detection comparative re- sults, F-measure and ACE value, on Arabic and Chinese data. The Arabic and Chinese models were built using 9In the English MaxEnt system, which uses 295k features, the distribution among the four classes of features is: 1:72%,... In PAGE 7: ... Unlike the English case, the systems had access to only a small amount of training data (60k words for Arabic and 90k characters for Chinese, in contrast with 340k words for English), which made it difficult to train statistical mod- els with large number of feature types. Future ACE evalu- ations will shed light on whether this lower performance, shown in Table3 , is due to lack of training data or to specific language-specific ambiguity. The final observation we want to make is that the sys- tems were not directly optimized for the ACE value, and there is no obvious way to do so.... In PAGE 7: ... The final observation we want to make is that the sys- tems were not directly optimized for the ACE value, and there is no obvious way to do so. As Table3 shows, the F-measure and ACE value do not correlate well: systems... ..."

"... In PAGE 31: ... This is also shown when greylisting agrees with the black lists. Table8 shows how many percent each real time black list marked as spam. A total of 340,841 messages was checked against the RBLs.... In PAGE 31: ... A total of 340,841 messages was checked against the RBLs. Since the graph in Figure 10 shows a higher number of spams detected when requiring one method than the black list that marked the most messages in Table8 , the black lists that mark the most messages as spam do not have exactly the same records in their database. 4.... ..."

"... In PAGE 7: ... In addition, we compare the spam detection algorithm with the substring amplification. We fix the set of 5 spams whose length and the number of spams are defined in Table2 . In a sample, 50 (resp.... ..."

"... In PAGE 15: ... SMART Net is a suitable tool for language teaching and learning. TABLE7 . continuetion Mean Scores 20.... ..."

"... In PAGE 5: ...Table1 ). The samples typically came from data posted recently on the Web, where the participants linguistic identity was not doubtful.... ..."

"... In PAGE 5: ... legitimate email In these experimens, a random 10% of the honeypot/cluster- spam data was used as queries against the honeypot/cluster- spam and the legitimate-email datasets. The resulting aver- age values of the recall are given in Table1 , which addition- ally shows relative increase in recall (see also Figure 3) due to signature randomization. None of the near-duplicate de- tection configurations produced any false-positive matches against the legitimate email collection.... In PAGE 5: ... This supports the claim that once a large diverse document collection is used, little in terms of copy-detection accuracy can be gained by tracking the changes to content distribution to fine tune the algorithm to the collection to which it is actually applied. The results shown in Figures 2 and 3, and in Table1 in- dicate that by using even a few extra randomized lexicons, the recall can be improved significantly. Given the rather small storage and computational consequences of using lex- icon randomization (linear in the number of lexicons used), this should make the proposed method attractive in practi- cal applications of I-Match.... ..."

"... In PAGE 3: ... This mixing between regular email users and spam senders can lead to more complex email networks than might have been naively expected and poses a challenging problem for detection. Table2 summarizes the number of addresses and emails by node classes and by internal or external origin. Node classes are as defined in Section 2 plus a third category - Spam amp; Non-Spam - which is the intersection of the former two.... ..."