Privacy protection of online social media users from malicious data miners

Sensitive information of Online Social Network (OSN) users can be discovered through sophisticated data mining, even if a user does not directly reveal such information. Malicious data miners can build a decision tree/s from a data set containing information about a huge number of OSN users and thereby learn general patterns which they can then use to discover the sensitive information of a target user who has not revealed the sensitive information directly. Therefore, data privacy is becoming increasingly important in the era of OSNs. The possibility of such a privacy breach has been empirically demonstrated in previous studies. An existing technique provides privacy by suppressing users' some non-sensitive attribute values (such as hometown) from their profiles. However, suppression of an attribute value may not be enough to secure a user's confidential information. In this study we experimentally demonstrate that (after taking necessary steps on non-sensitive attribute values) a user's sensitive information can still be inferred through his/her friendship information.

We propose three techniques in this study for protecting users' sensitive information from being inferred. All our proposed techniques are applied to a training data set to discover the general patterns. Based on the patterns, the proposed techniques can protect the sensitive information of the users in the testing data set.

Our first proposed protection technique, namely 3LP, considers both attribute values and friendship information while protecting a sensitive information for its users. Once the testing data set is modified following the suggestions made by 3LP we measure the data utility along with the privacy level to evaluate the effectiveness of the privacy techniques. Our second proposed technique, namely 3LP+, can protect users' multiple sensitive information. It takes a co-ordinated approach in each run while protecting sensitive attribute values. Our third technique is 3LPEx. Some previous techniques offer privacy protection based on specific classifiers. That is, while protecting privacy they use a specific classifier to learn the patterns and provide privacy based on these patterns. They assume that attackers will also use the same classifier and learn similar patterns. However, in reality, it is difficult to predict the classifier an attacker might use during a privacy attack. Therefore, 3LPEx, uses an exhaustive approach to learn the patterns instead of relying on a single classifier and uses the patterns to provide privacy.

We apply our proposed techniques on three OSN data sets. The experimental results show that our proposed methods outperform the existing privacy-preserving algorithms in terms of securing privacy while maintaining the data utility.