Source-linked AI summary
Combating Phone Scams with LLM-based Detection: Where Do We Stand?
Zitong Shen, Kangzhong Wang, Youqian Zhang, Grace Ngai, Eugene Y. Fu
TL;DR
Phone scams cause financial and emotional harm, and scammers continue to evade existing protections. This paper examines LLM-based detection through call-dialogue analysis; LLMs performed better than traditional classifiers on real-world scams, while low recall and hallucinations remain concerns.
Problem
Phone scams continue despite existing protections, and the paper examines whether LLMs can detect them by analyzing call dialogues.
Source: Introduction · p. 1Method
The paper evaluates LLM-based phone-scam detection by analyzing scam-call dialogues.
Source: Introduction · p. 1Results
LLMs demonstrated superior performance to traditional classifiers in identifying real-world scams, with Our-Real precision over 0.98 and recall as low as 0.72.
Source: Conclusion · p. 2, LLM-based Detector · p. 2Takeaways & Limitations
The paper concludes that scam detection must contend with the precision–recall trade-off and potential LLM hallucinations.
Source: Conclusion · p. 2On Our-Real, recall can be as low as 0.72, meaning 28 of 100 fraudulent calls could go undetected.
Source: Relatively Low Recall · p. 2, LLM-based Detector · p. 2
Abstract
from arXiv · showhide
Phone scams pose a significant threat to individuals and communities, causing substantial financial losses and emotional distress. Despite ongoing efforts to combat these scams, scammers continue to adapt and refine their tactics, making it imperative to explore innovative countermeasures. This research explores the potential of large language models (LLMs) to provide detection of fraudulent phone calls. By analyzing the conversational dynamics between scammers and victims, LLM-based detectors can identify potential scams as they occur, offering immediate protection to users. While such approaches demonstrate promising results, we also acknowledge the challenges of biased datasets, relatively low recall, and hallucinations that must be addressed for further advancement in this field
Introduction
Phone scams cause financial and emotional harm, while scammers adapt tactics to evade existing protections. LLMs may detect scams by analyzing call dialogues, but biased datasets, low recall, and hallucinations pose practical challenges.
- Scammers use spoofed numbers and deceptive tactics, including psychological manipulation and deepfakes, to evade protections and deceive victims.Source: Introduction · p. 1
- LLMs offer the prospect of detecting phone scams by analyzing call dialogues, but practical challenges remain.Source: Introduction · p. 1
- Biased datasets can make models overly sensitive to specific keywords and vulnerable to adversarial tactics.Source: Introduction · p. 1
- Low recall can leave scammers undetected, while hallucinations can hinder accurate detection in complex or ambiguous conversations and produce false positives or negatives.Source: Introduction · p. 1
Analysis of Existing Datasets
The analysis combines public scam datasets with authentic call transcripts and a newly generated dataset. Although classifiers score highly on existing data, the analysis finds that keyword reliance leaves them vulnerable when conversations change.
Data Collection
- Scam datasets included public transcripts labeled SC, SD, and MASC, authentic calls transcribed from public videos, and a separately generated Our-Synt dataset.Source: Data Collection · p. 1
Bias in Datasets
- TF-IDF features from the top 50 terms trained classic classifiers that exceeded 0.99, but success primarily reflected frequently occurring fraudulent-conversation terms.Source: Bias in Datasets · p. 1
- Table 1 presents Random Forest performance on the datasets.Source: Table 1 · p. 2
- Newly generated conversations containing the top 50 terms from real conversations caused a dramatic performance decline in previously trained classifiers.Source: Bias in Datasets · p. 2
LLM-based Detector
LLMs outperformed traditional classifiers in phone-scam detection, though their recall remained limited. The paper also evaluates response reliability, finding that it varies across models and datasets.
LLM-based Detector
- LLMs outperformed traditional classifiers, with Our-Real precision over 0.98 but recall as low as 0.72.On synthetic data, precision exceeded 0.95 and recall was as low as 0.86.Source: LLM-based Detector · p. 2
Relatively Low Recall
- A recall of 0.72 means 28 of every 100 fraudulent calls could go undetected, while higher recall must be balanced against precision.The authors say recall approaching 1 is desirable to minimize false negatives without misclassifying genuine conversations.Source: Relatively Low Recall · p. 2
Hallucination
- GPT-4o scored 1 on Our-Real, while GLM-4 scored 0.83 on Our-Synt, indicating model- and dataset-dependent reliability.Figure 1 presents performance for different LLM models.Source: Figure 1 · p. 2, Hallucination · p. 2
- The reliability score is normalized from 0 to 1, with higher values indicating greater reliability and fewer hallucinations.The score is based on discrete semantic entropy, and repeated runs revealed inconsistent classifications for some samples.Source: Hallucination · p. 2
Conclusion
Traditional classifiers performed well initially on biased datasets but remained vulnerable to adversarial attacks. LLMs performed better at identifying real-world scams, yet recall and hallucination challenges persist.
- Traditional machine-learning models achieved high initial performance on biased datasets but were susceptible to adversarial attacks.Source: Conclusion · p. 2
- LLMs demonstrated superior performance in identifying real-world scams.Source: Conclusion · p. 2
- The conclusion identifies precision–recall trade-offs and potential hallucinations as ongoing challenges requiring future work.Source: Conclusion · p. 2