| With the characteristics of fast sharing and easy-to-use, online social networks (OSNs) has become an inseparable part of our daily lives. Now it is routinely for users to post their recent experiences, communicate with their friends, keep track of the hottest trends and view interesting photos or videos through OSNs. On the other hand, OSNs also provides a new platform for attacks to gain profit. For example, attackers create fake accounts or compromise normal accounts (also named abnormal accounts) to launch malicious activities such as advertising, phishing and propagating pornography, which have seriously threatened both users’privacy and OSNs’reputation. In order to solve these problems, we do a series researches on spotting abnormal accounts in OSNs. Specifically, we focus on detecting a new type of spam which we named as Photo Spam. The main contributions of this paper can be summarized as follows:(1) We analyze the detection techniques on spotting abnormal accounts in OSNs. The life cycle of abnormal accounts consists of creation, development and application. Based on the distinct manifestations, abnormal accounts can be classified and unified in an architecture. The existing ways to spot abnormal accounts including data acquisition, data identification and result validation are also summarized. Finally, we introduce a new type of spam attack called photo spam that is able to evade the detection systems in OSNs. Compared to tradition spam, photo spam is more difficult to be detected and more harmful to normal users.(2) We propose a supervised learning solution to detect photo spammers in OSNs. There are two kinds of behaviors accounts in photo spam attacks. However, the existing schemes to detect photo spammers focus on only one kind of behavior. As a result, the other kind of behavior accounts cannot be detected. In our work, we first extract user-based and content-based features in photo spammers. Then we build a classification model with a dataset of 2,046 labeled users to identify photo spammers. Finally, we apply our detection model to 85,148 unlabeled users and detect 5,756 photo spammers with an accuracy rate of 97.05%.(3) We propose a novel lightweight iterative detection algorithm(LIDA) to detect photo s-pammers. In order to provide a high quality user experience, OSNs should detect spam accounts efficiently in a short period of time. However, the detection algorithms based on data mining need to traverse all accounts. In order to detect as many as spammers in a short time, we propose LIDA algorithm that contains target filter and content detection. In target filter, we identify more suspicious accounts according to the known spammers, and then we decide whether the suspicious accounts are really spammers by content detection. It is wor-thy that LIDA is a lightweight algorithm to infer more spam accounts through suspicious accounts instead of traversing all accounts. Experimental results in RenRen, which has suc-cessfully detected 9,568 spam accounts,30,732 spam albums and 2,626,780 spam URLs in four round iterations, indicate that LIDA is effective and efficiency in detecting photo spam accounts.(4) We propose a solution to detect spam albums in photo spam. Most of the approaches to detect photo spam are based on user malicious activities, which results in that spam accounts are detected after a significant lag time. In other words, spam accounts can only be detected after the malicious activities happens.To solve this problem, we propose a solution to detect spam albums. We first design 12 extracted easily and calculated efficiently features based on the differences between spam albums and normal albums. Next we build a classification model with a dataset of 2,356 labeled albums to identify spam albums. Our model provides excellent performance with true positive rates of spam albums and normal albums achieving 100% and 98.2% respectively. Finally, we apply our detection model to 315,115 unlabeled albums and detect 89,163 spam albums with an accuracy rate of 94.2%.
|
PDF Full Text Request