Download as PDF, PPTX
Uploaded byAmendra Shrestha
Identifying Warning Behaviors of Violent Lone Offenders in Written Communication
This document discusses using machine learning to identify warning behaviors in written communication that could indicate a violent lone offender (VLO). It presents experiments analyzing written texts from VLOs, bloggers, and forum users to classify them correctly with up to 92% accuracy. Key findings include that machine learning can separate VLO texts from others based on linguistic features and help identify potential VLOs prior to an attack. However, the document notes ethical issues would need consideration for real-world use.
Identifying Warning Behaviors of Violent Lone Offenders in Written Communication
- 1. Identifying Warning Behaviorsof Violent Lone Offenders in Written Communication 1Amendra Shrestha 1 Lisa Kaati 2 Tony Sardella 1Uppsala University 2Washington University December 12, 2016
- 2. Outline Introduction CounteringVLOs Data Experiments Conclusion 1 Introduction Example Violent lone offenders 2 Countering VLOs VLOs LIWC 3 Data 4 Experiments 5 Conclusion - 1 -
- 3. Outline Introduction CounteringVLOs Data Experiments Conclusion Example School shootings https://everytownresearch.org/school-shootings/ - 2 -
- 4. Outline Introduction CounteringVLOs Data Experiments Conclusion Example Lone actor terrorist attacks - 3 -
- 5. Outline Introduction CounteringVLOs Data Experiments Conclusion Example Mass murderers - 4 -
- 6. Outline Introduction CounteringVLOs Data Experiments Conclusion Violent lone offenders Violent Lone Offenders (VLO) • VLOs : school shooters, lone actor terrorists, mass murderers • wide factors : social status, ideology, mental health, personality type • rare events • pose a serious security threat to a society • shows sign of psychological warning behaviours • challenging to detect prior to an event • challenge to identify, target and arrest • common that they leave digital trace prior to attack - 5 -
- 7. Outline Introduction CounteringVLOs Data Experiments Conclusion Violent lone offenders Mass murderer : Dylan Roof • killed 9 persons in a church shooting in Charleston, South Carolina • published a manifesto on a website supporting white supremacy - 6 -
- 8. Outline Introduction CounteringVLOs Data Experiments Conclusion Violent lone offenders Lone actor terrorist: Anders Breivik • killed 8 people by detonating a van bomb in Oslo • shot dead 69 participants of a Workers’ Youth League • distributed a compendium of texts describing his militant ideology - 7 -
- 9. Outline Introduction CounteringVLOs Data Experiments Conclusion VLOs Countering VLOs • analyze and understanding potential signals in written communication • can be used to stop these attacks • combine weak signals and gain informations about intentions • weak signals • signs of an individuals radical beliefs and extreme hate • knowledge about how to produce homemade explosives • interest in firearms and signs of rehearsal • signs of warning behaviours from written text - 8 -
- 10. Outline Introduction CounteringVLOs Data Experiments Conclusion VLOs - 9 - Investigate possibilities to identify potential violent lone offenders based on written communication using machine learning
- 11. Outline Introduction CounteringVLOs Data Experiments Conclusion VLOs - 10 - • Electronic and written text (manifestos, letters, blogs, etc.) ⇓ comparision ←−−−−−→ Profile of VLOs text Profile of non-VLOs users text
- 12. Outline Introduction CounteringVLOs Data Experiments Conclusion LIWC LIWC - 11 - • Linguistic Inquiry and Word Count • a computerized word counting tool • counts words in psychologically meaningful categories
- 13. Outline Introduction CounteringVLOs Data Experiments Conclusion LIWC - 12 - Psychologist User
- 14. Outline Introduction CounteringVLOs Data Experiments Conclusion LIWC - 13 -
- 15. Outline Introduction CounteringVLOs Data Experiments Conclusion LIWC - 14 -
- 16. Outline Introduction CounteringVLOs Data Experiments Conclusion LIWC - 15 -
- 17. Outline Introduction CounteringVLOs Data Experiments Conclusion LIWC LIWC Categories - 16 -
- 18. Outline Introduction CounteringVLOs Data Experiments Conclusion Data - 17 - Figure : Jose Reyes’s Letters
- 19. Outline Introduction CounteringVLOs Data Experiments Conclusion Data • VLOs • manifesto, personal letter, suicide letter written by school shooters, mass murderers and lone offenders • 32 violent lone offenders : 46 documents • Non-VLOs • 54 blogs written about personal interests, news, fashion and photography • 108 stormfront users and their posts • 108 boards.ie users and their posts - 18 -
- 20. Outline Introduction CounteringVLOs Data Experiments Conclusion Experiment Setup • Feature selection : Mahalanobis relevance estimate • Synthetic Minority Over-sampling Technique (SMOTE) • Leave-One-Out Cross-Validation (LOOCV) • Adaboost • Java and R - 19 -
- 21. Outline Introduction CounteringVLOs Data Experiments Conclusion Experiment 1: Weak signals of warning behavior • if it is possible to separate texts written by VLO • combined lone offenders into one set • combined blogs, Stormfront and Boards.ie data into one set • 11 important features used • results : • Accuracy : 0.8766 • Blogs + Forums : 254 out of 270 are correctly classified • VLO : 33 out of 46 are correctly classified - 20 -
- 22. Outline Introduction CounteringVLOs Data Experiments Conclusion Experiment 2: Bloggers • possibility to identify lone offenders from bloggers • 12 important features used • results : blog vs VLO • Accuracy : 0.89 • Blogs : 50 out of 54 are correctly classified • VLO : 39 out of 46 are correctly classified - 21 -
- 23. Outline Introduction CounteringVLOs Data Experiments Conclusion Experiment 3: Stormfront users • identify lone offenders from Stormfront users • 10 important features used • results : Stormfront vs VLO • Accuracy : 0.9026 • Forum : 100 out of 108 are correctly classified • LO : 33 out of 35 are correctly classified - 22 -
- 24. Outline Introduction CounteringVLOs Data Experiments Conclusion Experiment 4: Boards.ie users • identify lone offenders from boards.ie users • 10 important features used • results : boards.ie vs VLO • Accuracy : 0.9221 • Forum : 100 out of 108 are correctly classified • LO : 42 out of 46 are correctly classified - 23 -
- 25. Outline Introduction CounteringVLOs Data Experiments Conclusion Conclusion • machine learning can be use to identify texts written by violent lone offenders • consider ethical issues • aid for human analyst - 24 -
- 26. Outline Introduction CounteringVLOs Data Experiments Conclusion - 25 - Thank You