Technical Summary: Domain Knowledge-Enhanced LLMs for Fraud and Concept Drift Detection

Problem Statement

Detecting deceptive conversations on dynamic digital platforms is increasingly challenging due to evolving language patterns and Concept Drift (CD). Concept drift refers to semantic or topical shifts that alter the context or intent of interactions over time. These shifts can obscure malicious intent or mimic normal dialogue, creating a significant hurdle for accurate classification.

While Large Language Models (LLMs) demonstrate strong performance in natural language tasks, they face critical limitations in high-stakes, risk-sensitive scenarios:

• Contextual Ambiguity: Difficulty distinguishing between genuine topic shifts (benign drift) and manipulative pivots (adversarial drift).
• Hallucinations: Tendency to generate incorrect information when context is ambiguous.
• Lack of Domain Adaptation: Standard LLMs often struggle to recognize nuanced patterns of deception without specific domain grounding.
• Static vs. Dynamic: Traditional machine learning (ML) models rely on fixed features and struggle to adapt to the fluid nature of multi-turn conversations where adversarial tactics evolve.

Existing approaches often treat fraud detection and concept drift modeling as isolated tasks, failing to account for their interdependence in real-world interactions.

Methodology

The authors propose a Domain Knowledge (DK)-Enhanced LLM framework that integrates pretrained LLMs with structured, task-specific insights. The methodology is executed in two distinct phases:

Phase I: Fake Review Detection (Static Settings)

• Objective: Validate the value of domain knowledge in a controlled, static environment.
• Dataset: The Yelp Fake Review Dataset (20,000 reviews: 15,000 real, 5,000 fake).
• Process:
  1. Pattern Discovery: The LLM analyzes pre-labeled examples to extract linguistic markers of fake reviews (e.g., exaggerated praise, vague sentiment, unnatural phrasing, inconsistent sentiment).
  2. Prompt Engineering: These identified patterns are embedded into structured prompts to guide the LLM's inference.
  3. Evaluation: Models (ChatGPT, DeepSeek, LLaMA, Claude) are tested under Zero-Shot (no domain guidance) and Domain-Knowledge (DK) conditions to measure the impact of structured prompts on classification accuracy.

Phase II: Fraud and Drift Detection in Streaming Dialogues

• Objective: Extend the framework to multi-turn conversations where deception evolves over time.
• Dataset: SEConvo (400 conversations: 24/191 fake, 16/169 real).
• Architecture: A Dual-LLM Framework comprising three interconnected components:
  1. DK-LLM Module 1 (Fraud Detection): Detects fake or deceptive conversations based on initial interaction patterns.
  2. OCDD (One-Class Concept Drift Detector): A statistical unit trained exclusively on non-malicious (normal) conversations using a One-Class Support Vector Machine (SVM). It monitors streaming dialogues to identify deviations (semantic shifts) from learned normal patterns.
  3. DK-LLM Module 2 (Drift Classification): If OCDD detects a drift, this second LLM interprets the shift. Using domain-specific cues (e.g., requests for PII, urgency, abrupt tone changes), it classifies the drift as either Benign (e.g., off-topic, spam) or Adversarial (e.g., phishing, manipulation).

Key Contributions

1. Domain Knowledge-Enhanced Dual-LLM Framework: A modular architecture that unifies fraud detection, concept drift identification, and drift classification. Unlike prior isolated approaches, this framework leverages the interdependence of these tasks.
2. Three-Stage Detection Pipeline: The system separates anomaly detection (DK-LLM), statistical drift detection (OCDD), and semantic interpretation (DK-LLM), enhancing robustness and interpretability in dynamic, multi-turn streams.
3. Impact of Structured Domain Prompts: The study demonstrates that embedding domain-specific knowledge into prompt design significantly boosts performance and explainability across both synthetic (fake reviews) and real-world (SEConvo) datasets.
4. Comprehensive Benchmarking: Extensive evaluation of multiple LLM backbones (ChatGPT, Claude, LLaMA, DeepSeek) under zero-shot and domain-guided settings, identifying optimal configurations for fraud and drift detection.

Results

• Phase I (Yelp Dataset): Incorporating domain knowledge significantly improved accuracy for most models.
  • Claude: Improved from 87% (Zero-Shot) to 95% (DK).
  • DeepSeek: Improved from 56% to 90%.
  • LLaMA: Improved from 52% to 58%.
  • ChatGPT: Showed marginal improvement (54% to 56%).
• Phase II (SEConvo Dataset): The DK-enhanced Dual-LLM framework significantly outperformed both zero-shot baselines and traditional ensemble models (kNN, SVM, Imp-Kmeans).
  • LLaMA achieved the highest performance, reaching 98% across Accuracy, Precision, Recall, and F1-Score with domain knowledge, compared to 90% in the baseline.
  • DeepSeek saw a notable gain in recall (67% to 88%), achieving an F1-score of 80%.
  • The DK-enhanced LLaMA model (98% accuracy) significantly outperformed the traditional ensemble model (82% accuracy).

Significance and Claims

The paper claims that integrating domain-specific knowledge into LLM pipelines is essential for high-stakes NLP applications. The primary significance lies in:

• Enhanced Accuracy and Robustness: Domain-guided prompting allows models to distinguish between genuine topical shifts and manipulative intent, reducing false positives and missed detections.
• Improved Interpretability: By separating statistical drift detection from semantic classification, the system provides clearer reasoning for its decisions, addressing the "black box" nature of standard LLMs.
• Adaptability: The framework addresses the limitations of static ML models in dynamic environments where adversarial tactics and language patterns continuously evolve.

The authors note that while LLaMA-based implementations showed superior performance (likely due to training on dialogue-centric corpora), the approach is generalizable. They acknowledge limitations, such as the system's current struggle with sarcasm, cultural idioms, and emotionally ambivalent language, and suggest future work in human-in-the-loop feedback and multilingual adaptation.