Privacy preserving data mining
Medical, financial, or social databases are analyzed daily for the discovery of pat- terns and useful information. Privacy concerns have emerged as some database segments contain sensitive data. Data mining techniques are used to parse, process, and manage enormous amounts of data while ensuring the...
Αποθηκεύτηκε σε:
| Κύριος συγγραφέας: | |
|---|---|
| Άλλοι συγγραφείς: | |
| Γλώσσα: | English |
| Δημοσίευση: |
2021
|
| Θέματα: | |
| Διαθέσιμο Online: | http://hdl.handle.net/11610/21812 |
| Ετικέτες: |
Προσθήκη ετικέτας
Δεν υπάρχουν, Καταχωρήστε ετικέτα πρώτοι!
|
| Περίληψη: | Medical, financial, or social databases are analyzed daily for the discovery of pat- terns and useful information. Privacy concerns have emerged as some database segments contain sensitive data. Data mining techniques are used to parse, process, and manage enormous amounts of data while ensuring the preservation of private information, as data can be exploited by potential aggressors. Regarding social networks, their privacy preserving analysis aims to understand better the network and its behavior, while at the same time protecting the privacy and identity of its individuals. Network data contain sensitive information and due to the increasing popularity of social networks that are released publicly, effective anonymization techniques are required to make the data available for research.
Considering the above, this thesis is divided in two parts and focuses on privacy preservation of distributed databases and social network data. In the first part, a privacy preserving data mining protocol is presented, thoroughly designed and developed for both horizontally and vertically partitioned databases, which contain either nominal or numeric attribute values. At the same time the accuracy of final outcomes and the preservation of privacy is the main goal of the proposed protocol. Cryptography, as shown by previous research, is the most accurate approach to acquiring knowledge while maintaining privacy to assure both confidentiality and integrity of data. The proposed algorithm exploits the multi-candidate election schema to construct a privacy-preserving tree-augmented naive Bayesian classifier, a more robust variation of the classical naive Bayes classifier. The exploitation of the Paillier cryptosystem and the distinctive homomorphic primitive shows in the security analysis that privacy is ensured and the proposed algorithm provides strong defences against common attacks.
In the second part, an anonymization algorithm is developed for weighted graphs, i.e., for social networks where the strengths of links are important. Previous studies concentrate mainly on preventing identity disclosure in unweighted graphs. How- ever, a weighted graph is more descriptive, revealing more information about the relationships between entities, which allows adversaries to take advantage of potential security holes. Weights can be essential for social network analysis, but they pose new challenges to privacy preserving network analysis. For instance, an adversary may use his information about some edge weights to re-identify individuals. This in contrast with many previous studies which only consider unweighted graphs. The proposed anonymization method considers identity, edge and edge weight disclosure for anonymizing weighted graph data, assuming that adversaries have knowledge about the neighborhood of a targeting entity. In particular, a k-anonymous technique is presented that groups entities with same neighborhoods into supernodes and the corresponding connections into superedges. The method provides k-anonymity of nodes against attacks where the adversary has information about the structure of the network, including its edge weights.
Both approaches are proven efficient and have been evaluated in terms of privacy and utility. Experiments deriving the benefits of real world databases demonstrate the preservation of private data while mining processes occur. |
|---|