Page 590 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

(continued)

Item Details
Model Training Anomaly Detection: Train models on NGO financial data to spot
and Fine- unusual transactions. Fine-tune with new data to catch evolving
fraud.
Tuning
Graph Analytics: Build networks of NGOs and related entities to
find suspicious connections (e.g., duplicate registrations). Adjust
detection parameters as needed.
Sentiment Analysis: Train NLP models on text data (social media,
news) to detect negative sentiment or fraud allegations. Update
with new text for better accuracy.
Geospatial & Satellite Data Analysis: Correlate reported project
locations with satellite images to verify existence. Refine image
analysis for better results.
OCR: Extract text from documents for digital analysis and
cross-checking.
Testbeds or Deployment is planned for Q3 2025 with select NGOs registered
Pilot Deployments on NGO Darpan [1], focusing on financial transaction monitoring
in Maharashtra. Proof of Concept (PoC) using synthetic data:

PoC Description: A PoC will use synthetic NGO financial data
(e.g., 1,000 transactions across 50 NGOs) and registration
records to test anomaly detection and duplicate registration
algorithms. The PoC will simulate fraud scenarios (e.g., duplicate
Permanent Account Numbers (PANs), unusual fund transfers)
and evaluate model accuracy.

Dataset: Generate synthetic data mimicking NGO Darpan’s
structure (financial transactions, registration details).

Expected Outcome: Achieve >85% accuracy in detecting fraudu-
lent transactions.

2 Use Case Description

2�1� Description

In India, Non-Governmental Organizations (NGOs) play a crucial role in social development.
However, the sector is vulnerable to various forms of fraud, including misuse of funds, duplicate
registrations, and the operation of shell NGOs solely to siphon government and Corporate
Social Responsibility (CSR) funds. These fraudulent activities undermine the effectiveness of
genuine NGOs, lead to significant financial losses, and erode public trust. Traditional manual
methods of scrutiny are often time-consuming, resourceintensive, and prone to human error,
making it difficult to effectively identify and prevent sophisticated fraud.

554