Research map / Large language models / Language, safety and interpretability
ScamNet: Toward Explainable Large Language Model-Based Fraudulent Shopping Website Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Fine-tunes Llama-3-8B-Instruct with two-step LoRA SFT (detection data then 200 curated explanations) to detect fraudulent shopping websites with explanations, outperforming BEYOND PHISH on a newer real-world dataset.
Problem
Detecting fraudulent shopping websites at scale is hard: prior feature-engineered classifiers are brittle to evolving scam tactics and lack explainability, and out-of-the-box LLMs underperform.
Approach
Two-step LoRA supervised fine-tuning of Llama-3-8B-Instruct on website URL, content, external links and WHOIS info, followed by a second SFT round on a small curated set of GPT-4-drafted and human-corrected explanations so the model outputs a label with justification.
Key results
- On the real-world dataset, SCAMNET reaches 95.59% TPR and 96.15% F-1 versus 22.35% TPR for BEYOND PHISH.
- On the BEYOND PHISH test set, SCAMNET achieves 99.47% accuracy and 0.22% FPR, versus 94.57% accuracy and 3.71% FPR for BEYOND PHISH.
- Human experts rated explanation quality at an average of 1.8 on a 0-2 scale.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fine-tune Llama-3-8B-Instruct with SCAMNET's two-step process instead of using BEYOND PHISH | Fraudulent shopping website detection · Real-World | TPR: 22.35% -> 95.59% | Helped |
| Fine-tune Llama-3-8B-Instruct with SCAMNET | Fraudulent shopping website detection · BP evaluation dataset | F-1: 90.19% -> 99.40% | Helped |
| Use GPT-4 2-shot prompting instead of 0-shot prompting | GPT-4 · BEYOND PHISH evaluation dataset | FPR: 11.23% -> 1.50% | Helped |
| Use two-step fine-tuning including the curated explainable dataset rather than step 1 only (ablation) | SCAMNET · Real-World | F-1: 0.47 -> 0.87 (rank 8); 0.65 -> 0.92 (rank 16); 0.53 -> 0.96 (rank 32) | Helped |
| Increase LoRA rank from 8 to 16 during step 1 (ablation) | SCAMNET · Real-World | F-1: 0.47 -> 0.65 | Helped |
| Increase LoRA rank from 16 to 32 during step 1 (ablation) | SCAMNET · Real-World | F-1: 0.65 -> 0.53 | Hurt |
| Increase LoRA rank from 16 to 32 during step 2 (ablation) | SCAMNET · Real-World | F-1: 0.92 -> 0.96 | Helped |
Surprising findings
- Out-of-the-box GPT-4 and GPT-3.5-turbo underperform the specialized BEYOND PHISH feed-forward classifier, and few-shot prompting gives little gain.
- Non-tuned Llama-3-8B-Instruct labels almost all websites as fraudulent.
Limitations
- Explanation quality assessed only by three experts on 200 websites with a coarse 0-2 scale.
- Real-world evaluation set is small on the fraud side (340 fraudulent websites) and drawn from one two-month window.
- Website content is truncated to fit the token limit.
Experimental setup
Datasets: BEYOND PHISH training and evaluation datasets explainable reasoning dataset (200 curated samples)
Benchmarks: BEYOND PHISH evaluation dataset Real-World dataset (collected May-June 2024)
Metrics: accuracy TPR FPR F-1
Baselines: BEYOND PHISH Llama-3-8B-Instruct GPT-3.5-turbo GPT-4
Builds on: LoRA BEYOND PHISH Llama-3-8B-Instruct
Closest papers by meaning
- Combating Phone Scams with LLM-based Detection: Where Do We Stand? (Student Abstract) (AAAI 2025 · similarity 0.88)
- On the Importance of Application-Grounded Experimental Design for Evaluating Explainable ML Methods (AAAI 2024 · similarity 0.83)
- T-NET: Weakly Supervised Graph Learning for Combatting Human Trafficking (AAAI 2024 · similarity 0.83)
- Bad Actor, Good Advisor: Exploring the Role of Large Language Models in Fake News Detection (AAAI 2024 · similarity 0.83)
- PhishAgent: A Robust Multimodal Agent for Phishing Webpage Detection (AAAI 2025 · similarity 0.82)
- Online Fraud Detection via Test-Time Retrieval-Based Representation Enrichment (AAAI 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.