StageGuard: A hierarchical BERT-BiLSTM framework with stage transition anomaly scoring for fraudulent conversation detection

Abstract

Fraud conducted through chat applications has grown into a serious global threat affecting millions of people each year. Most detection systems still classify each message in isolation rather than reasoning over the conversation as a whole, which limits their ability to recognise the gradual manipulation that characterises social-engineering fraud.

This paper presents StageGuard, a hierarchical deep-learning architecture that models the full trajectory of a fraudulent conversation. The framework builds on the Manipulation Stage Progression Model (MSPM), a five-stage taxonomy grounded in Cialdini's principles of psychological influence [4], in which offenders build trust gradually across turns before exploiting it. On this basis we define the Stage Transition Anomaly Score (STAS), a single-valued, human-readable feature that quantifies how abruptly a conversation moves between stages and serves as an interpretable indicator of manipulative intent.

BERT-base-uncased encodes each turn, and the resulting turn embeddings are processed by a bidirectional LSTM [6], [19] with Bahdanau attention [7] that captures temporal dependencies across the conversation. The attention context vector is fused with the STAS and urgency features before a multi-layer classifier produces the fraud decision, and training uses a multi-task objective that jointly optimises fraud detection and per-turn stage classification. Under a leakage-free protocol in which messages are partitioned into disjoint train, validation and test pools before conversations are constructed, and on a realistic task in which both classes contain spam so that content alone is insufficient, StageGuard attains an F1-macro of 0.9919 and an ROC-AUC of 0.9998. The hierarchical design substantially outperforms flat baselines that treat the conversation as a single text (F1 up to 0.952); ablation and McNemar tests indicate that STAS does not change accuracy significantly, and its value is instead the per-turn interpretability it provides for human operators. We further report bootstrap confidence intervals, five-fold cross-validation and an external-domain evaluation that exposes a sharp generalisation gap, which we discuss candidly as a limitation.