Secure Messaging Analysis

Overview

This project created a privacy-preserving analysis tool for organizations working with sensitive communications from at-risk communities.

The Challenge

Human rights organizations need to understand communication patterns and sentiment trends among communities they serve, but accessing message content directly poses serious privacy and security risks to vulnerable individuals.

The Solution

I developed an on-device analysis system that processes communications locally and only shares aggregate, anonymized insights:

On-Device Processing: All text analysis happens on the user's device. No raw message content ever leaves the device.

Differential Privacy: Aggregate statistics are protected with differential privacy guarantees, ensuring individual communications cannot be reconstructed.

Pattern Analysis: The system identifies trends in communication volume, sentiment shifts, and topic clustering without accessing specific messages.

Alerting: Organizations can set thresholds for unusual patterns that might indicate a crisis or security concern.

Technical Approach

The system uses lightweight NLP models (DistilBERT fine-tuned for the domain) that run efficiently on mobile devices. Federated learning techniques allow the models to improve over time without centralizing data.

Impact

The tool has been adopted by two organizations working with refugee communities, helping them identify emerging needs and potential security concerns while maintaining the trust and privacy of the people they serve.