Abstract
Cyberstalking represents a growing social challenge characterized by repeated, personalized harassment across digital platforms. Traditional spam filters often fail to capture the longitudinal "course of conduct" or the highly contextual nature of such threats. This paper introduces the Anti Cyberstalking Text-based System (ACTS), a multi-modular machine learning framework designed to identify cyberstalking incidents and automate the collection of forensic evidence. By integrating stylometrics, user-defined personalization, and cryptographic documentation, ACTS aims to bridge the gap between automated detection and legal prosecution.
Introduction
Cyberstalking is a pervasive and escalating social problem, yet it remains notoriously difficult to combat. Unlike isolated phishing attempts or generic spam, cyberstalking is defined by a "course of actions"—repeated incidents that cause significant distress to the victim. Current defensive technologies face three major hurdles: the Definition Gap (difficulty tracking longitudinal patterns), the Personalization Paradox (context-dependent threats), and the Burden of Proof (the technical difficulty of preserving admissible evidence). To address these, the ACTS framework is proposed as a comprehensive solution for emails, SMS, and Instant Messages (IMs).
The ACTS Framework
The ACTS framework moves beyond simple keyword blocking by utilizing five specialized modules that work in concert to evaluate incoming messages and manage the response workflow.
- Detection Module: Employs standard NLP techniques such as tokenization and stemming alongside supervised learning models (e.g., Neural Networks or SVMs) to categorize content as cyberstalking, legitimate, or "grey."
- Attacker Identification Module: Uses stylometrics and Principal Component Analysis (PCA) to create "writeprints." These lexical and structural profiles help link anonymous or pseudonymous messages to a single perpetrator.
- Personalisation Module: A "human-in-the-loop" component allowing victims to define specific rules and maintain a "code dictionary" of sentiment-heavy words relevant to their unique situation.
- Aggregator Module: Acts as the system's decision engine, taking numerical outputs ($\alpha, \beta, \lambda$) from the other modules and applying a decision function ($\psi$) to determine the message's final classification.
- Messages and Evidence Collection Module: Automatically gathers forensic data, including source IPs and headers, securing them with hash functions and asymmetric encryption to ensure legal admissibility.
Key Insights and Design Philosophy
A central tenet of the ACTS framework is that context is king. The authors emphasize that cyberstalking is often highly personalized; for example, a message as innocent as "Happy Birthday" can be a terrifying threat if sent by a stalker to a victim who recently suffered a loss. By including a Personalisation Module, the system acknowledges that machine learning alone cannot fully grasp the specific trauma-triggers of a victim without user-defined parameters.
Furthermore, the introduction of a "grey" category prevents the pitfalls of binary classification. This allows the system to flag suspicious messages for further review rather than simply deleting them or letting them through.
Crucially, ACTS treats documentation as being as important as detection. Integrating cryptographic hashing into the detection pipeline ensures that the "chain of custody" for digital evidence begins the moment a message is received, significantly reducing the emotional and technical burden on the victim during potential legal proceedings.
Implementation and Limitations
The current work describes a prototypical implementation focused on architectural requirements. While the system is designed to run on a user's device to collect real-world, personalized data, several challenges remain:
- The Short-Text Problem: Stylometrics (writeprints) are difficult to apply to short-form content like SMS or Tweets due to the limited linguistic signal available in 160 characters.
- Cold Start for Identification: The Attacker Identification module requires a history of messages ($n \ge 2$) to function. The system cannot easily identify a stalker on the very first message if they have no prior history in the database.
- Evaluation Pending: As this is a framework proposal, the primary next step is full implementation and empirical evaluation against real-world harassment scenarios to measure its effectiveness and refine its predictive accuracy.
Conclusion
The ACTS framework provides a logical proof-of-concept for aggregating diverse machine learning techniques into a single engine capable of addressing the complexities of cyberstalking. By prioritizing both detection and forensic documentation, it offers a more robust path toward victim protection and legal recourse than traditional filtering methods.
Summary updated