Research map / Large language models / Language, safety and interpretability
2-in-1 Phishing Detection via Large LM Distillation and Small LM Perturbation (Student Abstract)
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.
Proposes DualLM, which distills GPT-3.5-Turbo rationales into a frozen small LM (BERT/RoBERTa/DistilBERT) and trains only a head plus trainable input perturbation tokens for phishing email detection with few labeled examples.
Problem
Phishing email detection with minimal labeled data, where fine-tuning small LMs or LLMs requires substantial data and many trainable parameters.
Approach
Combines chain-of-thought rationale distillation from a large LLM (aligning the small LM's [cls] representation to rationale embeddings via an MSE loss) with a frozen small LM that is adapted only through trainable perturbation tokens appended to the input and a classification head.
Key results
- DualLM-BERT reaches 74.87% F1 on Phishing-Type vs. 61.37% for BERT-H and 14.88% for BERT-F.
- DualLM-DistilBERT reaches 97.31% F1 on Phishing-Fraud vs. 92.92% for WARP and 92.54% for Distilling Step-by-Step.
- DualLM-RoBERTa reaches 94.04% F1 on Phishing-Spam vs. 91.09% for WARP; the paper claims up to 10,000x fewer trainable parameters.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add rationale guidance and trainable input perturbations to the small LM | DualLM · Phishing-Type, Phishing-Spam, and Phishing-Fraud | F1 score (%) | Helped |
| Increase the number of training shots (ablation) | DualLM · Phishing-Type, Phishing-Spam, and Phishing-Fraud | Accuracy | Helped |
| Increase the number of input perturbations (ablation) | DualLM · Phishing-Type, Phishing-Spam, and Phishing-Fraud | Accuracy | Mixed |
| Add rationale guidance (ablation) | BERT baseline · Phishing-Type, Phishing-Spam, and Phishing-Fraud | Improvement percentage | Helped |
| Add trainable input perturbations (ablation) | BERT baseline · Phishing-Type, Phishing-Spam, and Phishing-Fraud | Improvement percentage | Helped |
Surprising findings
- Frozen-backbone head-only training (BERT-H, DistilBERT-H) outperforms full fine-tuning (BERT-F, RoBERTa-F) on the small Phishing-Type dataset in the reported table.
Limitations
- Small evaluation scale: Phishing-Type has only 160 emails, and results are reported with a fixed n=10 shots.
- Performance does not monotonically improve with more perturbation tokens; the paper notes larger t does not always help.
- Rationales depend on a proprietary LLM (GPT-3.5-Turbo) and the paper does not analyze rationale quality.
Experimental setup
Datasets: Phishing-Type Phishing-Spam Phishing-Fraud
Benchmarks: Phishing-Type Phishing-Spam Phishing-Fraud
Metrics: F1 score
Baselines: BERT-F RoBERTa-F DistilBERT-F BERT-H RoBERTa-H DistilBERT-H Distilling Step-by-Step WARP
Builds on: Distilling Step-by-Step WARP chain-of-thought prompting
Closest papers by meaning
- Teach LLMs to Phish: Stealing Private Information from Language Models (ICLR 2024 · similarity 0.84)
- Bad Actor, Good Advisor: Exploring the Role of Large Language Models in Fake News Detection (AAAI 2024 · similarity 0.84)
- Privacy Auditing of Large Language Models (ICLR 2025 · similarity 0.83)
- MIA-Tuner: Adapting Large Language Models as Pre-training Text Detector (AAAI 2025 · similarity 0.83)
- Local Sparsity Enables Unsupervised LLM Safety Detection (NeurIPS 2026 · similarity 0.83)
- Combating Data Laundering in LLM Training (NeurIPS 2026 · similarity 0.83)
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.