Indian Job Fraud Detector

A machine learning system that detects fraudulent job postings in the Indian market - deployed as a live API, Chrome extension, and public web interface.

Live demo: https://aditsawhney.github.io/FraudGuard/

---

Overview

Job fraud is a significant problem in India. Scams involving upfront registration fees, Aadhaar/PAN collection, WhatsApp-only contacts, and MNC impersonation are widespread. No publicly labeled dataset of Indian job fraud existed, so this project builds one from scratch and trains a model on it.

The system combines TF-IDF text analysis with 15 hand-engineered India-specific fraud signals, wrapped in a FastAPI backend and surfaced through a Chrome extension that runs inline on LinkedIn job pages.

---

Model Performance

Metric	Score
Test accuracy	0.94
Fraud F1	0.94
Real F1	0.93
False positives	8
False negatives	12

Evaluated on a held-out test set of 318 rows.

Caveat: the real class is majority synthetic. These metrics likely overstate real-world performance - no evaluation on genuine LinkedIn postings has been done yet.

Evaluation charts

---

Dataset

Three-source hybrid — 2,118 rows total.

Source	Rows	Label
Groq-generated synthetic Indian jobs	452	real
Groq-generated synthetic Indian jobs	300	fraud
EMSCAD benchmark dataset	500	real
EMSCAD benchmark dataset	866	fraud

The 300 synthetic fraud rows cover 6 Indian-specific archetypes (50 each): fee-based scams, data harvesting, guaranteed placement, simple task scams, MNC impersonation, and overseas visa fraud. Generated using data_pipeline/generate_dataset.py, grounded in a documented taxonomy in data_pipeline/indian_fraud_taxonomy.py.

EMSCAD adds real human-written Western scam postings to diversify the fraud class beyond purely synthetic data.

EMSCAD citation: "Employment Scam Aegean Dataset", University of the Aegean, 2014. Kaggle

Synthetic Indian dataset: Kaggle — Indian Job Fraud Dataset

---

Model

Architecture: CombinedFeatureExtractor — a single sklearn transformer that computes:

• TF-IDF (1–2 ngrams, 30k features, sublinear_tf)
• 15 hand-crafted Indian fraud regex features
• 2 meta-features: total_flags, flag_density

These are hstacked into one sparse matrix and fed into a LogisticRegression classifier (class_weight='balanced', C=1.0).

Logistic regression was chosen deliberately: coef_ gives direct feature-level explainability, inference is fast enough for Render's free tier, and the dataset (2,118 rows) is too small to justify deep models.

15 Indian fraud signals

fee_language · upfront_payment · unrealistic_salary · guaranteed_job · whatsapp_only · gmail_contact · personal_phone · aadhaar_request · pan_request · bank_details_early · mnc_impersonation · background_check_fee · overseas_visa_fee · simple_task_scam · vague_company

---

Project structure

fraud-job-detector/
├── website/                       ← public-facing site (GitHub Pages)
│   ├── index.html
│   ├── css/
│   ├── screens/
│   ├── js/
│   └── images/
├── data/
│   ├── raw/
│   │   ├── synthetic_indian_jobs.csv
│   │   └── emscad_fake_job_postings.csv
│   └── processed/
│       ├── train.csv / val.csv / test.csv
│       └── dataset_metadata.json
├── data_pipeline/
│   ├── indian_fraud_taxonomy.py   ← 6 fraud archetypes, documented
│   ├── generate_dataset.py        ← synthetic data generation
│   └── merge_datasets.py          ← combines synthetic + EMSCAD
├── model/
│   ├── features.py                ← CombinedFeatureExtractor
│   ├── train.py
│   ├── model.pkl
│   ├── evaluation_report.json
│   └── top_features.json
├── backend/
│   ├── app.py                     ← FastAPI, POST /predict
│   └── requirements.txt
├── extension/
│   ├── manifest.json
│   ├── content.js
│   ├── background.js
│   ├── popup.html / popup.js / popup.css
│   └── sites/
│       └── linkedin.js            ← DOM scraper
└── reports/
    ├── confusion_matrix_test.png
    ├── confusion_matrix_val.png
    ├── feature_coefficients.png
    ├── precision_recall_f1.png
    └── dataset_composition.png

---

Running locally

Backend

cd backend
pip install -r requirements.txt
uvicorn app:app --reload

The backend runs at http://127.0.0.1:8000. model/features.py must be present alongside app.py for the pickle to deserialise correctly.

Website

cd website
python3 -m http.server 8080

Open http://localhost:8080.

Extension

1. Go to chrome://extensions
2. Enable Developer Mode
3. Click "Load unpacked" → select the extension/ folder
4. Navigate to a LinkedIn job posting on linkedin.com/jobs/view/

---

API

POST /predict — https://fraud-job-detector-1mdz.onrender.com

Request:

{
  "title": "Software Engineer",
  "company": "Acme Corp",
  "location": "Bangalore",
  "salary": "₹8-12 LPA",
  "contact": "",
  "description": "..."
}

Response:

{
  "prediction": "fraud",
  "confidence": 0.91,
  "risk_level": "High",
  "reasons": ["aadhaar number requested upfront", "whatsapp-only contact"],
  "top_model_features": [
    {"feature": "indian_aadhaar_request", "coefficient": 1.04},
    {"feature": "indian_whatsapp_only", "coefficient": 0.74}
  ]
}

Full docs: https://fraud-job-detector-1mdz.onrender.com/docs

---

Known limitations

• Metrics are measured on a partially synthetic dataset and likely overstate real-world performance.
• Short legitimate company names (EY, TCS) occasionally trip the vague_company feature.
• LinkedIn's search results layout (/jobs/search-results/) uses obfuscated class names and is not supported. The extension prompts users to open the full job view instead.
• Render free tier spins down after inactivity - first request after idle takes ~30 seconds. The extension and website fall back to local heuristics during this time.

---

Author

Built by Adit Sawhney