An AI-powered real-time detection system that identifies phishing and social engineering attacks using BERT + Psychological Influence Analysis.
Social engineering attacks are responsible for 90% of all cyber breaches worldwide. Traditional rule-based systems fail against modern AI-generated phishing attacks.
SEAD-AI solves this by combining:
- 🧠 Fine-tuned BERT — detects language manipulation patterns
- ⚡ PsychoGuard Engine — detects Cialdini's 6 Influence Principles
- 🔍 Explainable AI — tells users exactly WHY a message is suspicious
- 📊 Real-time scoring — combined threat score in milliseconds
The first known system to encode Cialdini's 6 Principles of Influence as machine learning features:
| Principle | What It Detects |
|---|---|
| Authority | Fake CEO/IRS/Bank impersonation |
| Urgency | "Act now or account deleted" |
| Scarcity | "Only 3 spots left!" |
| Social Proof | "Millions of users trust us" |
| Reciprocity | Fake gifts to create obligation |
| Liking | False rapport and flattery |
Coming soon on Render.com
Paste any message → Get instant threat analysis
- Threat Score meter (0-100%)
- Psychology triggers detected
- Explainable AI reasons
- Recommended action
| Metric | Sample Data | Full Dataset |
|---|---|---|
| Accuracy | 83.33% | 99.05% |
| Precision | 75.00% | 99.05% |
| Recall | 100.00% | 99.05% |
| F1 Score | 85.71% | 99.05% |
| Dataset Size | 30 rows | 23,467 rows |
| Component | Technology |
|---|---|
| AI Model | BERT (bert-base-uncased) |
| Framework | PyTorch + HuggingFace |
| Backend | Flask (Python) |
| Database | SQLite |
| Frontend | HTML + CSS + JavaScript |
| GPU | NVIDIA CUDA |
git clone https://github.com/YOUR_USERNAME/sead-ai.git
cd sead-aipython -m venv venv
# Windows
venv\Scripts\activate
# Mac/Linux
source venv/bin/activatepip install -r requirements.txtpython datasets/download_datasets.pypython preprocessing/clean.pypython models/train_bert.pypython app.pyOpen: http://localhost:5000
docker-compose up --buildOpen: http://localhost:5000
python tests/test_all.pyExpected:
| Method | Endpoint | Description |
|---|---|---|
| POST | /analyze | Analyze a message |
| GET | /history | Scan history |
| GET | /stats | Dashboard stats |
| GET | /search?q= | Search history |
| GET | /health | System health |
import requests
response = requests.post("http://localhost:5000/analyze",
json={"text": "URGENT: Your account is suspended!"})
print(response.json())
# {
# "threat_score": 78,
# "risk_level": "High Risk",
# "reasons": ["Urgency tactics detected", ...]
# }| Dataset | Size | Type |
|---|---|---|
| CEAS 2008 | 32,204 rows | Phishing |
| Enron Emails | 15,000 rows | Legitimate |
| Nazario Corpus | 5,000 rows | Phishing |
| Nigerian Fraud | 5,000 rows | Phishing |
| SpamAssassin | 6,000 rows | Mixed |
Total: 23,467 training samples
- Fakhouri et al. — AI-Driven Solutions for Social Engineering (IEEE, 2024)
- Wang et al. — Adversarial AI Detection (IJST, 2025)
- Cialdini, R. — Influence: The Psychology of Persuasion (1984)
- Thorne — Cognitive Vulnerabilities in the Age of LLMs (TAJET, 2025)
B.Tech Cybersecurity Research Project Built as part of IBM Skills project — 2025 by Himang and Devendra
MIT License — free to use and modify.