One Size Fits All? A Modular Adaptive Sanitization Kit (MASK) for Customizable Privacy-Preserving Phone Scam Detection

TL;DR

This work addresses the privacy risks of cloud-based LLMs in phone scam detection by introducing MASK, a modular, trainable sanitization kit that enables user-driven privacy controls via a pluggable sanitization stack. The framework comprises a privacy preference adapter, a modular sanitization layer with an extensible plugin architecture, and four initial strategies (TF-IDF, PII statistical, PII masking, and transcript summarization) to balance privacy and detection accuracy. Through experiments on real Chinese scam data with multiple LLMs, the authors show that PII Masking Anonymization provides the best privacy-utility balance, achieving complete PII removal with minimal impact on detection performance, while more aggressive abstraction strategies reduce semantic retention and accuracy. The work argues for a configurable, adaptive approach to privacy in LLM-based detection and highlights potential for on-device, privacy-preserving implementations and future trainable sanitizers that jointly optimize privacy, semantics, and accuracy.

Abstract

Phone scams remain a pervasive threat to both personal safety and financial security worldwide. Recent advances in large language models (LLMs) have demonstrated strong potential in detecting fraudulent behavior by analyzing transcribed phone conversations. However, these capabilities introduce notable privacy risks, as such conversations frequently contain sensitive personal information that may be exposed to third-party service providers during processing. In this work, we explore how to harness LLMs for phone scam detection while preserving user privacy. We propose MASK (Modular Adaptive Sanitization Kit), a trainable and extensible framework that enables dynamic privacy adjustment based on individual preferences. MASK provides a pluggable architecture that accommodates diverse sanitization methods - from traditional keyword-based techniques for high-privacy users to sophisticated neural approaches for those prioritizing accuracy. We also discuss potential modeling approaches and loss function designs for future development, enabling the creation of truly personalized, privacy-aware LLM-based detection systems that balance user trust and detection effectiveness, even beyond phone scam context.

Figures (5)

• Figure 1: LLM-based phone scam detection approaches. Cloud-based models possibly risk privacy leakage, while data sanitization (at user/local devices) enables safe detection with cloud models.
• Figure 2: Overview of privacy-preserving frameworks: (a) Existing methods employ fixed (e.g., PII-like entity masking) with one-size-fits-all approaches, offering limited flexibility and no user preference customization. (b) The proposed Modular Adaptive Sanitization Kit (MASK) enables dynamic privacy adjustment through a pluggable architecture that accommodates various sanitization methods such as traditional keyword-based and PII-like methods for users with different privacy preferences. It also provides extensible plugin capabilities for future enhancements, particularly for sophisticated neural network approaches.
• Figure 3: Overview of the four sanitization strategies: (1) TF-IDF keyword representation, (2) PII statistical representation, (3) PII masking anonymization, and (4) transcript summarization.
• Figure 4: Trade-off between PII removal rate (PRR), semantic retention rate (SRR), and detection performance. (a) Circle size represents precision. (b) Circle size represents recall. Colors indicate different sanitization strategies.
• Figure 5: Trainable sanitization framework that jointly optimizes privacy, semantic retention, and task accuracy.