"... In PAGE 45: ... The problem with a real-world application is that these privacy-preserving data mining techniques have not yet been adopted on a large scale. Table5 presents some examples found in the literature in which the methods were evaluated with real data. It can be clearly seen that four of the seven studies presented used synthetic data instead of real data.... In PAGE 45: ... It can be clearly seen that four of the seven studies presented used synthetic data instead of real data. The use of real data does not produce better results; the idea that almost none of the methods is really adapted to real-world applications can be seen from Table5 . As can also be seen in Table 5, various data mining algorithms have been considered in isolation from each other.... ..."

"... In PAGE 28: ... The main objective in privacy-preserving data mining is to develop algorithms for modifying the original data in some way, so that the private data and private knowledge remain private even after the mining process [1]. This approach requires definition what is private, and, because this has many different definitions (examples can be seen in Table3 ), the approaches ... In PAGE 30: .... does not constitute an intrusion. Formal definitions for both of these are challenging. In Table3 , some examples of how privacy is said to be preserved in the literature are presented. To the best of my understanding, only Agrawal et al.... In PAGE 76: ...g. [61]), while a few others try to formulate privacy within a mathematical framework (see, for example, Table3 ). The privacy issue is controversial and, as the discussion in Chapter 2.... ..."

"... In PAGE 64: ... Although the descriptive information might provide interesting insights about the data, the actual use of the clustering method for predictions will be the cluster assignments The cluster assignments tell us which segment each customer belongs to and, in case of a new customer, the potential segment. Table6 shows the assignments of the models, how the assignments of the protocols differ. The assignments of the normal k-means are taken as the baseline to which the assignments of our privacy-preserving k-means are being compared.... In PAGE 64: ... The assignments of the normal k-means are taken as the baseline to which the assignments of our privacy-preserving k-means are being compared. The results in Table6 show that misassignments grow as the number ... ..."

"... In PAGE 61: ... Although the descriptive information might provide interesting insights about the data, the actual use of clustering method on predictions will be the cluster assignments The cluster assignments tell us which segment each customer belongs to and in case of a new customer, the potential segment. The Table6 shows the assignments of the models, how the assignments of the protocols differ. The assignments of the normal k means are taken as baseline to which the assignments of our privacy preserving k means are being compared at.... In PAGE 61: ... The assignments of the normal k means are taken as baseline to which the assignments of our privacy preserving k means are being compared at. The results in the Table6 shows that the miss assignments grow as the number of... ..."

"... In PAGE 14: ...3 Results Our first battery of tests broadly profile the performance of OPE and DPE. Shown in Table1 , the most striking characteristic of these experiments is that they demonstrate that OPE protocols consume two orders of magnitude more network resources than the DPE protocols. These costs can be directly attributed to the oblivious transfer algo- rithms whose primitive cryptographic operations require the transfer of many polyno- mials between hosts.... ..."

"... In PAGE 12: ...3 Results Our first battery of tests broadly profile the performance of OPE and DPE. Shown in Table1 , the most striking characteristic of these experiments is that they demonstrate that OPE protocols consume two orders of magnitude more network resources than the DPE protocols. These costs can be directly attributed to... ..."

"... In PAGE 1: ... Anomaly detection systems [8, 16, 4, 20] can be used to score these flows, and the analyst typically looks at only the most anomalous flows to identify attacks or other undesirable behavior. Table2 shows 17 flows which were ranked as most suspicious by the MINDS Anomaly Detec- tion Module [8] for the network traffic analyzed on January 26, 2003 (48 hours after the Slammer Worm hit the Inter- net). These flows are involved in three anomalous activities - slammer worm related traffic on port 1434, flows asso- 1Traditionally, a centroid is defined as the average of the value of each attribute over all transactions.... In PAGE 2: ... If many of these most anomalous flows can be summarized into a small representation, then the analyst can analyze a much larger set of anomalies than is otherwise possible. For ex- ample, if the dataset shown in Table2 can be automatically summarized into the form shown in Table 3 (the last col- umn has been removed since all the transactions contained the same value for it in Table 2), then the analyst can look at only 3 lines to get a sense of what is happening in 17 flows. Table 3 shows the output summary for this dataset generated by an application of our proposed scheme.... In PAGE 2: ... If many of these most anomalous flows can be summarized into a small representation, then the analyst can analyze a much larger set of anomalies than is otherwise possible. For ex- ample, if the dataset shown in Table 2 can be automatically summarized into the form shown in Table 3 (the last col- umn has been removed since all the transactions contained the same value for it in Table2 ), then the analyst can look at only 3 lines to get a sense of what is happening in 17 flows. Table 3 shows the output summary for this dataset generated by an application of our proposed scheme.... ..."