Supraja_SMS_presentation
0 likes158 views
This document describes research conducted on detecting fake Twitter accounts by analyzing profile characteristics rather than tweet content. Over 62 million Twitter profiles were collected and grouped using attributes like name, description, location, and patterns in screen names. An entropy-based approach was used to identify profiles with similar screen name patterns that may belong to the same group or person, in order to detect automatically generated or synthetic accounts at scale. The goal is to help identify fake or inauthentic accounts that could bias analyses of Twitter data.
Supraja_SMS_presentation
- 2. Fake Twi)er accounts: Profile characteris6cs obtained using an ac6vity-‐based pa)ern detec6on approach Supraja Gurajala, Joshua S. White, Brian Hudson, and Jeanna N. Ma:hews Department of Mathema?cs and Computer Science Clarkson University, Potsdam, NY gurajas@clarkson.edu
- 3. Online Social Networks • OSNs are widely used for: – Personal communica?ons, social interac?ons – Commercial uses • Marke?ng through OSNs • Brand growth – Big Data • E.g. Elec?on data analysis • Flu tracking – Our focus -‐ Twi:er 2
- 4. Mining Twi:er Data • New and novel applica?ons – Public health – Personal /Organiza?onal popularity – Scien?fic dissemina?on • Necessitates confidence in user authen?city – Challenged by cyber-‐opportunists • Sybils – fake profiles created to mimic users – Some general, some are Iden?ty clone accounts 3
- 5. Fake Accounts -‐ problem • “Followers for hire” – Infla?ng follower numbers for other accounts • Fake followers – Bias an individual/organiza?on’s popularity – Alter characteris?cs of the audience – Create a legi?macy problem • A mul?-‐million dollar market for buying fake followers 4
- 6. Proposed approach • Most of the current Twi:er-‐based research has focused on analysis of tweets • Twi:er profile-‐based analysis – Faster detec?on – Simpler logis?cs • smaller storage, memory required – More comprehensive detec?on • Can also detect accounts that don’t tweet 5
- 7. Implementa?on • Twi:er user profiles – Created accounts to access profiles – Twi:er API – Breadth first search over a set of seed users – Currently collected 62 million user profiles – JSON (JAVA Script Object Nota?on) format • Mongodb 6
- 8. Methodology • 33 a:ributes for each profile • Manual inspec?on – Most of these a:ributes are set to default values • 10 Key a:ributes – name, followers_count, friends_count, verified, created_at, descrip?on, loca?on, updated, profile_image_url and screen_name 7
- 9. Raw User Profiles – 62 Million 21.2 Million accounts 3.8 Million groups NameGroup by Methodology
- 10. Raw User Profiles – 62 Million 21.2 Million accounts 3.8 Million groups Name e.g. Group name: FREE FOLLOW Number of groups: 1 Number of accounts :1833 Group by Methodology
- 11. Raw User Profiles – 62 Million 21.2 Million accounts 3.8 Million groups Name Name + Description + Location e.g. Group name: FREE FOLLOW Number of groups: 1 Number of accounts :1833 6.9 Million accounts 0.7 Million groups Methodology Group by
- 12. Raw User Profiles – 62 Million 21.2 Million accounts 3.8 Million groups Name e.g. Group name: FREE FOLLOW Number of groups: 1 Number of accounts :1833 6.9 Million accounts 0.7 Million groups Number of groups: 78 Number of accounts: 804 Name + Description + Location Methodology Group by
- 13. Raw User Profiles – 62 Million 21.2 Million accounts 3.8 Million groups Name e.g. Group name: FREE FOLLOW Number of groups: 1 Number of accounts :1833 6.9 Million accounts 0.7 Million groups Number of groups: 78 Number of accounts :804 Screen-name Patterns Name + Description + Location Methodology Group by
- 14. Pa:ern Recogni?on 13 Group
- 15. Pa:ern Recogni?on 14 Group
- 16. Pa:ern Recogni?on 15 Group Name1 Base Screen_Name Calculate SE1 SE – Shannon Entropy
- 17. Pa:ern Recogni?on 16 Name2 Select Screen-‐Name for analysis SE – Shannon Entropy Group Name1 Base Screen_Name Calculate SE1
- 18. Pa:ern Recogni?on 17 Name2 Concatenate screen-‐names Select Screen-‐Name for analysis SE – Shannon Entropy Group Name1 Base Screen_Name Calculate SE1 Name1Name2
- 19. Pa:ern Recogni?on 18 Name2 Concatenate screen-‐names Select Screen-‐Name for analysis SE – Shannon Entropy Group Name1 Base Screen_Name Calculate SE1 Name1Name2 SE2
- 20. Pa:ern Recogni?on 19 Name2 Concatenate screen-‐names Select Screen-‐Name for analysis SE – Shannon Entropy Group Name1 Base Screen_Name Calculate SE1 Name1Name2 SE2 Is SE2-‐SE1<0.1
- 21. Pa:ern Recogni?on 20 Name2 Concatenate screen-‐names Select Screen-‐Name for analysis SE – Shannon Entropy Group Name1 Base Screen_Name Calculate SE1 Name1Name2 SE2 Is SE2-‐SE1<0.1 Add Name2 to current collec?on Yes
- 22. Pa:ern Recogni?on 21 Name2 Concatenate screen-‐names Select Screen-‐Name for analysis SE – Shannon Entropy Group Name1 Base Screen_Name Calculate SE1 Name1Name2 SE2 Is SE2-‐SE1<0.1 Add Name2 to current collec?on Yes No Add to group for analysis in next itera?on
- 23. Pa:ern Recogni?on 22 Name2 Concatenate screen-‐names Select Screen-‐Name for analysis SE – Shannon Entropy Group Name1 Base Screen_Name Calculate SE1 Name1Name2 SE2 Is SE2-‐SE1<0.1 Add Name2 to current collec?on Yes No Add to group for analysis in next itera?on
- 24. Pa:ern Recogni?on 23 Name2 Concatenate screen-‐names Select Screen-‐Name for analysis SE – Shannon Entropy Group Name1 Base Screen_Name Calculate SE1 Name1Name2 SE2 Is SE2-‐SE1<0.1 Add Name2 to current collec?on Yes No Add to group for analysis in next itera?on
- 25. Pa:ern Recogni?on 24 Name2 Concatenate screen-‐names Select Screen-‐Name for analysis Group Name1 Base Screen_Name Calculate SE1 Name1Name2 SE2 Is SE2-‐SE1<0.1 Yes No Add to group for analysis in next itera?on
- 26. Pa:ern Recogni?on 25 Name2 Concatenate screen-‐names Select Screen-‐Name for analysis Group Name1 Base Screen_Name Calculate SE1 Name1Name2 SE2 Is SE2-‐SE1<0.1 Yes No Add to group for analysis in next itera?on
- 27. Raw User Profiles – 62 Million 21.2 Million accounts 3.8 Million groups Name e.g. Group name: FREE FOLLOW Number of groups: 1 Number of accounts :1833 6.9 Million accounts 0.7 Million groups 8276 patterns Accounts: 263,250 Number of groups: 78 Number of accounts :804 Screen-name Patterns e.g. Patterns in a collection list: freefollow, SavedCompete, etc. Name + Description + Location Methodology Group by
- 28. Elimina?on of false posi?ves • Iden?fies closely associated accounts • False posi?ves – Highly popular names • Final filter – Determine distribu?on of update ?mes • Genuine accounts – uniform distribu?on • Fake -‐ non uniform 27
- 29. 28 Update (me occurrences Chopra free_follow Distribu?on of update ?mes as a func?on of ?me of day
- 30. Raw User Profiles – 62 Million 21.2 Million accounts 3.8 Million groups Name e.g. Group name: FREE FOLLOW Number of groups: 1 Number of accounts :1833 6.9 Million accounts 0.7 Million groups 8276 patterns Accounts: 263,250 8873 groups Accounts: ~56,000 Number of groups: 78 Number of accounts :804 e.g. Patterns in a collection list: freefollow, SavedCompete, etc. Update times Core group of fake accounts: Same Name, Description, Location, and Update times and Screen names matching 1 or more patterns Screen-name Patterns Name + Description + Location Methodology Group by
- 31. Analysis of Results • Genera?on of Ground Truth data – Random sample of Twi:er user profiles – Similar size and ?meline for consistency • Fake Profile characteris?cs – Update Times – Crea?on Times – URL Analysis 30
- 32. Update ?mes 31 Ground truth Fake Profiles
- 33. Update days 32 Ground truth Fake profiles
- 34. Crea?on ?mes 33 Ground truth Fake Profiles No bias in the crea?on ?mes of ground truth data, unlike the fake profiles
- 35. Crea?on days 34 Ground truth Fake profiles
- 36. Crea?on ?me distribu?on • Fake accounts are mostly created in batches, over short ?me intervals • Note: these results are for ~60000 accounts 35
- 37. Crea?on ?me distribu?on • Fake accounts are mostly created in batches, over short ?me intervals • Note: these results are for ~60000 accounts 36
- 38. Crea?on ?me distribu?on • Fake accounts are mostly created in batches, over short ?me intervals • Note: these results are for ~60000 accounts 37
- 39. 38 Collage of images from dis?nct URLs for a group of 659 accounts within a fake profile set
- 40. Conclusions • A large Twi:er profile database was analyzed – Fake accounts were detected based on their profile characteris?cs • Fake accounts tend to have minimal diversity in: – crea?on ?mes – update ?mes – image URLs • The crea?on of fake accounts is possibly only semi-‐automated 39
- 41. Future Work • Using this core fake data as seed will help us iden?fy a more comprehensive fake profile set • Pa:ern-‐matching focused Algorithm • Understand the intra and inter rela?onship of these fake accounts. • Tweet Analysis -‐ Apache SPARK -‐ GraphX 40