TruScamBuster: Federated Latent-Thought Models for Zero-Shot Scam SMS Detection on Mobile Devices with Lightweight LLMs
Scam short message service (SMS) remains a persistent global risk, increasingly leveraging obfuscation—homoglyph substitutions, leetspeak, and code-switching—to evade detectors. Conventional spam filters and token-level classifiers degrade under such shifts, while large language models are too costly for phones. We introduce TruScamBuster, a lightweight, privacy-preserving framework that performs latent-thought prototype inference on device and uses federated learning (FL) to adapt without sharing raw SMS. Compact encoders (DistilBERT, MobileBERT, TinyLlama-1.1B) map messages to a latent space in which classification reduces to nearest-prototype comparison; FL periodically aggregates encoder updates and class prototypes to sustain adaptability. Across public benchmarks and a crowdsourced scam corpus, TruScamBuster yields consistent in-domain gains over token-level baselines, delivers larger improvements in zero-shot settings on unseen variants (e.g., homoglyphs and English–Chinese code-switching), and converges faster and more stably in federated simulations. These results indicate that coupling prototype-based latent reasoning with FL provides a practical path to robust, efficient, and privacy-preserving scam SMS detection on mobile devices.
Paper
Full text
TruScamBuster: Federated Latent-Thought Models for Zero-Shot Scam SMS Detection on Mobile Devices with Lightweight LLMs
Semantic Scholar · 2025
Abstract
Scam short message service (SMS) remains a persistent global risk, increasingly leveraging obfuscation—homoglyph substitutions, leetspeak, and code-switching—to evade detectors. Conventional spam filters and token-level classifiers degrade under such shifts, while large language models are too costly for phones. We introduce TruScamBuster, a lightweight, privacy-preserving framework that performs latent-thought prototype inference on device and uses federated learning (FL) to adapt without sharing raw SMS. Compact encoders (DistilBERT, MobileBERT, TinyLlama-1.1B) map messages to a latent space in which classification reduces to nearest-prototype comparison; FL periodically aggregates encoder updates and class prototypes to sustain adaptability. Across public benchmarks and a crowdsourced scam corpus, TruScamBuster yields consistent in-domain gains over token-level baselines, delivers larger improvements in zero-shot settings on unseen variants (e.g., homoglyphs and English–Chinese code-switching), and converges faster and more stably in federated simulations. These results indicate that coupling prototype-based latent reasoning with FL provides a practical path to robust, efficient, and privacy-preserving scam SMS detection on mobile devices.