Abstract
Ensuring the safety of juveniles in online social networks is a critical component of public safety. However, the sheer volume of online conversations makes manual monitoring impossible, necessitating automated, scalable detection techniques. This research addresses two primary challenges in Online Predator Identification (OPI): the difficulty of obtaining representative "normal" conversation data for training, and the performance limitations of traditional machine learning models. We propose a dual-method approach: utilizing semi-supervised anomaly detection (One-class SVM) to identify predatory behavior as a deviation from the norm, and a customized Convolutional Neural Network (CNN) to improve classification accuracy. Results on the PAN-2012 and Sûreté du Québec (SQ) datasets demonstrate that these methods significantly enhance detection capabilities, with the CNN achieving a 1.7% improvement in F1-score over traditional Support Vector Machines.
Introduction
Online Predator Identification (OPI) is a high-stakes domain where the cost of a false negative can be devastating. Despite its importance, the field faces significant technical hurdles. Traditional binary classification requires balanced datasets of "predatory" and "normal" behavior. However, defining "normal" is an open-ended problem, as everyday conversation spans an infinite range of topics. This paper explores how anomaly detection and deep learning can bypass these limitations to create more robust and field-ready detection systems.
The "Negative Sample" Problem and Anomaly Detection
A major bottleneck in OPI research is the scarcity of representative non-predatory data. Most benchmark datasets use technical logs (e.g., IRC chat) as "normal" samples, which fail to capture the nuances of general social interaction. This leads to models that are effective at spotting technical jargon but poor at identifying predators in general-purpose chat rooms.
To address this, we treat predatory behavior as an anomaly. By employing a One-class SVM, the model can be trained primarily on known predatory instances. This approach is particularly valuable for law enforcement agencies that possess archives of criminal logs but lack curated, "clean" datasets of everyday conversations. Our experiments show that with proper noise removal—such as filtering out single-participant "conversations"—the F1-score for anomaly detection can reach 75%, proving its viability as a semi-supervised solution.
Deep Learning for Nuanced Classification
While anomaly detection provides a path forward for imbalanced data, improving absolute classification performance requires more sophisticated architectures. We customized a Convolutional Neural Network (CNN) for text analysis. Unlike standard NLP tasks that might use pre-trained models like Word2Vec, this approach builds internal word embeddings during training. This preserves the specific "slang," intentional misspellings, and grooming language unique to predators—features that general-purpose models often smooth over.
The CNN's hierarchical structure allows it to extract abstract features and local patterns (n-grams) across long chat logs. In empirical tests using the PAN-2012 dataset, this customized CNN outperformed traditional baselines, including standard SVMs and Artificial Neural Networks, by approximately 1.7% in F1-score.
The Importance of Stylistic Signatures
A key insight from this research is the counter-intuitive role of text preprocessing. In many NLP applications, stemming and stop-word removal are standard procedures to reduce dimensionality. However, in OPI, these techniques can be detrimental. Predators often use specific stylistic markers—such as elongated words ("soryyyyyyyyy") or unique punctuation—to build rapport or mimic the speech patterns of children. Preserving these "stylistic fingerprints" is essential for the model to distinguish between a genuine juvenile and a predator attempting to blend in.
Real-World Validation: The SQ Dataset
To test the practical utility of these models, we applied them to a dataset provided by the Sûreté du Québec (SQ). This dataset consists of French-language conversations gathered from actual criminal investigations. While the sample size was small (82 conversations), it served as a crucial "proof of concept" for the "Resolute" software prototype, demonstrating that the proposed methods can be adapted to real-world, multi-lingual forensic environments.
Conclusion and Future Directions
This research demonstrates that combining anomaly detection with deep learning offers a powerful toolkit for identifying online predators. By treating predation as a statistical outlier and leveraging CNNs to capture context-specific language, we can overcome the data scarcity and performance plateaus of earlier methods. Future work will explore Social Network Analysis (SNA) to identify predator cliques and the use of Paragraph2Vec to further enhance the semantic understanding of grooming behaviors within digital environments.
Summary updated