GenAI Phishing Detector
Fine-tuned DistilBERT for three-way classification of financial-sector messages in the Nigerian context.
| Label | Class | Description |
|---|---|---|
| 0 | Legitimate | Authentic bank communications (debit alerts, transfer confirmations, KYC reminders) |
| 1 | Traditional Phishing | Low-effort phishing with grammatical errors and suspicious links |
| 2 | AI-Generated Phishing | Polished, professionally written LLM-generated phishing |
Usage
from transformers import pipeline
classifier = pipeline(
"text-classification",
model="Mirianuli/genaimitproject",
truncation=True,
max_length=128,
)
print(classifier("URGENT!!! Your BVN has been BLOCKED. Click here to verify now: https://account-verify.tk/38472"))
Training data
5000 procedurally generated samples (seed 42) across three classes.
Evaluation caveat
The original random-split evaluation reports 1.0000 across all metrics. That
split allows template-sibling samples to appear in both train and test, so the
score overstates real-world performance. A template-aware group split with zero
shared template families is included in the repository for an honest comparison;
see experiments/template_aware_split/ and notebooks/template_aware_experiment.ipynb.
Framework versions
- transformers
- PyTorch