Development of a lightweight interpretable model for an early-stage detection of child grooming indicators in short conversational text.

Online child grooming is a dangerous, gradual process where predators build trust with minors to facilitate abuse and prevent detection. While various software tools exist to combat this threat, the vast majority rely on post-hoc analysis. This means they review chat logs only after the grooming has occurred or escalated, failing to stop immediate danger. To address this critical gap, this research aims to design and implement a lightweight, client-side framework that enables real-time, message-by-message identification to interrupt predatory tactics as they happen.

A major hurdle for real-time detection is the widespread use of end-to-end encryption (E2EE) in modern messaging apps, which blocks platforms and law enforcement from viewing conversations. To bypass this barrier without compromising core security protocols, this study utilizes client-side scanning (CSS). This architecture analyzes conversation data directly on the child's device before it is encrypted and transmitted. While previous image-based scanning proposals faced intense public pushback over mass surveillance concerns, this research focuses strictly on text-based natural language processing (NLP) classifiers, which carry a much lower privacy risk.

Beyond encryption, current detection models face severe technical limitations regarding how they analyze data. Grooming rarely happens in a single exchange; it unfolds subtly over time. Most existing tools only analyze isolated messages or short conversational windows, meaning they lose track of early warning signs. This study introduces a dedicated context-preservation mechanism that retains relevant conversation history locally, allowing the system to accurately identify long-term patterns of manipulation and isolation.

The effectiveness of current tools is further restricted by outdated and limited data. Most available datasets are over a decade old and rely entirely on English, failing to account for modern slang, transliterations, and multilingual interactions. Because online grooming is a global issue, this framework seeks to establish a more reliable foundation for detection. It balances performance and accuracy, ensuring that the local software can successfully catch grooming behaviors while keeping false positives low and minimizing the impact on the device's battery and processing power.

Ultimately, this study addresses a critical spike in online exploitation, which has expanded across social media, private messaging apps, and online multiplayer games. Predators frequently use these spaces to make initial contact before moving children to encrypted platforms. By restricting the software to devices used by minors under informed parental consent, this framework establishes a strong legal and ethical justification. It provides a necessary, proportional solution that protects vulnerable children from severe harm while respecting broader digital privacy rights.