Downloaded 27 times
More Related Content
Similar to Master's Thesis Presentation
Master's Thesis Presentation
- 1. Proposal of aTerrorist Detection Model in Social Networks Master’s thesis defense Presented By : Wajdi Khattel on 07.12.2019 2018 / 2019 In front of jury composed of: ● President: Najet AROUS ● Evaluator: Olfa EL MOURALI ● Academic supervisor: Ramzi GUETARI ● Laboratory supervisor: Nour El Houda BEN CHAABENE
- 2. Outline 2 1 Introduction 2 3 Proposed Model 5 Implementation & Results 6 Conclusion &Perspective Existing works
- 3.
- 4. Context ▰The appearance ofsocial networks created an ease of communication ▰The usage of social networks differs: Friendly vs harmful ▰Terrorists are one of the most dangerous category ▰The detection of these users is important 4
- 5. Problematic ▰Terrorists tend tohide their abnormal behavior ▰Normal user could adopt terrorist behavior ▰Socio-cultural definition of a terrorist could change over time ⇒ Time is important 5
- 6. Objective ▰Propose a terroristdetection model ▻Consider over-time user’s behavior change ▻Consider over-time behavior’s definition change ▰Cover Limitation of existing models 6
- 7.
- 8.
- 9. Anomaly Detection 9 Paper Input Format DescriptionMultiple Social Networks Multiple Input data types User’s Behavior Change Behavior Definition Change Lashakry et al., 2019 Activity Proposal of model for user profile creation to monitor users ✓ ✓ ✗ ✗ Zamanian et al.,2019 Activity Proposal of model for user activity pattern recognition ✗ ✓ ✓ ✗ Bhattacharjee et al., 2017 Graph Proposal of a probabilistic anomaly classifier mode ✗ ✗ ✓ ✓ Chen et al., 2018 Graph Proposal of a user profiling framework that can be used to detect anomalous users ✗ ✓ ✗ ✗
- 10. Proposed Input Format 10 HybridInput Format: ▰Graph Input ▰Activity-based score for each node
- 11. 11 Terrorism Detection Alvari etal. (2019) - Different data collecting methods - Textual-content data features Chitrakar et al. (2016) Kalpakis et al. (2019) - Advantages of using Convolutional Neural Network (CNN) - Advantages of using Transfer Learning Technique - Advantages of using multidimensional networks - Social Network Analysis methodologies
- 12.
- 13. ▰Model Input: MultidimensionalNetwork ▰Three sub-models: ▻Text classification model ▻Image classification model ▻General Information classification model ▰Decision Making 13 Proposed Model
- 14. Model Input 14 ▰Nodes: Users ▰Dimensions:User’s social medias content ▰Edges: Connection between users on a certain dimension Multidimensional Network
- 15. ▰Input: Textual data ▰Process: ▻NaturalLanguage Processing ▻Word Embedding ▻Machine Learning classification ▰Output: Score 15 Text Classification Model (TCM)
- 16. ▰Objective: Make themachine able to understand the human language ▰Process: ▻Morphological Analysis ▻Syntactical Analysis ▻Semantical Analysis 16 TCM: Natural Language Processing
- 17. ▰Objective: Represent textin a numerical way while preserving its semantics ▰Process: ▻Term Frequency-Inverse Document Frequency (TF-IDF) 17 TCM: Word Embedding
- 18. ▰Input: Image data ▰Process: ▻Usepre-trained convolutional neural network model ▻Add new convolutional layers ▰Output: Score 19 Image Classification Model (ICM)
- 19.
- 20. ▰Input: General Informationdata ▰Process: ▻If data is non-numerical ⇒ Encode it ▻Machine Learning classification ▰Output: Score 21 General Information Classification Model
- 21. ▰Input: 3 submodelsscores ▰Process: ▻Calculate user score ▻Classify it based on threshold ▰Output: User category (Terrorist or not) 22 Decision Making TCM ICM GICM Decision Making S1 S1 = Score1 * Weight1 S2 = Score2 * Weight2 S3 = Score3 * Weight3 S2 S3
- 22.
- 23.
- 24. ▰Offline Data: Dataused for the model training ▻Textual Data: Tweets from banned Twitter accounts ▻Image Data: Images from google image ▻General Information Data: PIRUS dataset ▰Online Data: Data used for testing and live usage ▻Facebook Graph API ▻Instagram REST API ▻Twitter REST API 25 Data Collection
- 25. 26 TCM: NLP +Word Embedding
- 26. 27 Label Number ofsamples Positive labels 122619 Negative labels 181691 Total Data 304310 TCM: Training Data
- 27. 28 Model Name AccuracyF1-Score Training Time Logistic Regression 0.9726 0.9674 39.9 secs SVM 0.9626 0.9548 6h 48min 33secs Neural Network 0.9774 0.9719 1min 11secs TCM: Classification Model
- 28. 29 Label Number ofsamples Positive labels 219 Negative labels 314 Total Data 533 ICM: Training Data
- 29.
- 30. 31 ICM: Classification Model ModelName Accuracy F1-Score Training Time CNN 0.7631 0.7219 3mins 50secs CNN + DA 0.7781 0.7463 4mins 12secs CNN + TL 0.8291 0.8103 8mins 48secs CNN + DA + TL 0.8571 0.8454 9min 23secs
- 31. 32 Label Number ofsamples Positive labels 114 Negative labels 126 Total Data 240 GICM: Training Data
- 32. 33 GICM: Classification Model ModelName Accuracy F1-Score Training Time Logistic Regression 0.7650 0.7873 5 secs SVM 0.8300 0.8495 7 secs Neural Network 0.8173 0.8325 48.6 secs
- 33. 34 Proposed Model ▰Text ClassificationModel: Neural Network ▰Image Classification Model: CNN + DA + TL ▰General Information Model: SVM
- 34.
- 35.
- 36. Conclusion 37 ▰Proof-of-concept of terroristdetection model ▰Working with multiple social networks and multiple data types ▰Supports over-time behavior change
- 37. Perspective 38 ▰Graph Analysis ▰Support moredata types: Video ▰Train on more data
- 38. 39 Thank you foryour attention !