1. Pay Me and I’ll Follow You: Detection of Crowdturfing Following Activities in Microblog Environment
Liu Yuli
2016/05/22

2. Background
The number of followers becomes an important metric of the influence and reputation of a person’s or a business entity’s account [Cha et al., 2010]
Some users turn to the follower market to get undeserved followers
The abnormal following activities impose a great threat
disrupt fair following mechanisms
help customers to gain excessive reputation or influence
malicious entities can also spread malwares and/or perform other spamming activities
1、在微博环境中 一个用户的粉丝数量是
2、正常来说，如果一个用户想要增加粉丝数量，就要让其它用户对他感兴趣，但这种方式是很困难的，所以有些用户不满足于这种正常关注，所以找到
5/5/2019

3. Background A new rise of malicious following activities
Fake Follower suggested by previous works
fraudulent accounts (i.e., fake accounts or Sybils)
created and manipulated by market operators for conducting spamming activities [Almaatouq et al., 2014; Motoyama et al., 2011]
compromised accounts
owned by legitimate users whose credentials have been stolen by spammers [Stringhini et al., 2012] to conduct spamming activities against their will
A new rise of malicious following activities
Voluntary follower （volower）
oriented from normal users
willing to join the follower market and get rewards
5/5/2019

4. Voluntary Follower Activities
Properties comparison between volowers and legitimate users
volowers are much more similar to legitimate users
voluntary follower accounts are owned by real users
be willing to follow customer
在我们的观测周期内
5/5/2019

5. Data Preparation Labeled user Unlabeled User (Uu)
Volower (voluntary follower Uv)
3,000 volowers purchased from follower market
Legitimate User (Un)
thirty thousand account
a sample set of legitimate users and their neighbours
Unlabeled User (Uu)
millions of accounts
all the labeled users’ neighbors
这里我们没有提到 customer账号
5/5/2019

6. DetectVC Algorithm Two assumptions
a customer will be followed by many volowers to gain influence or reputation
a volower will follow many customers to gain enough profit
5/5/2019

7. DetectVC Algorithm
Detect voluntary followers and customers simultaneously
seed set Us  Uv is a small set of users that are randomly selected from labeled volowers Uv
each user u will receive two scores Pv(u) and Pc(u) that indicate its possibilities of being a volower and a customer
介绍检测算法的实现细节
5/5/2019

8. DetectVC Algorithm Using the form of matrix/vector
=W T = W
An example of DetectVC calculation
5/5/2019

9. Experiments Performance of customer detection seed volowers
randomly select the seed set with size m for 100 times
performance
rank all the users in descending order of Pc(u)
randomly sample 2,000 users and manually label
1、说明种子的有效性
2、检测算法的稳定性
首先设定一定的种子数量 m 从100到1000 间隔为100 对于每个种子数量m 我们随机的选择100 也就是做100次实验
5/5/2019

10. Experiments Performance of volower detection
ranked the users in our user set in descending order of Pv(u), and use the top 3,000 users to show the effectiveness
detect a user u as volower if Pv(u) > r
5/5/2019

11. Experiments Comparison with Baseline Methods
comparison of F-measure scores between our volower detection method and baselines
comparison of F-measure scores between our customer detection method and baselines
Original
With Pv(u)
DetectVC
0.844 
[Yang et al., 2012]
0.715
0.850 (+13.5%)
[Egele et al., 2013]
0.807
0.863 (+5.6%)
[Lee et al., 2014]
0.832
0.895 (+6.3%)
[Aggarwal et al., 2015]
0.825
0.868 (+4.3%)
Original
With Pc(u)
DetectVC
0.860 
[Stringhini et al., 2013]
0.805
0.864 (+5.9%)
[Aggarwal et al., 2015]
0.837
0.907 (+7.0%)
5/5/2019

12. Conclusion
Investigated the new problem of voluntary following activity detection
DetectVC incorporates prior knowledge and graph structure
Does not involve any static or dynamic property features
5/5/2019

13. Thanks