Lie Detection System

Usage

Me and 2 of my uni colleagues made this project for AI in medicine subject because we thought the given topics were too easy. (It's just a proof of concept version. Maybe someday I'll find some time to improve it)

This project combines hardware and software to create a system that can detect lies by analyzing physiological signals. Here's a quick overview of what this project entails and how everything fits together.

Overview

This project develops a lie-detection system using a combination of hardware and software. The system is integrated with an app that displays questions and records the answers.

Example of obtained signals

Components

Hardware

• Arduino UNO
• Pulse oximeter sensor (Infrared + Red LEDs)
• Galvanic skin resistance sensor

Software

• Desktop Application:
  • Collects training data by asking questions and prompting the person whether they have lied.

Hardware Setup

The sensors are connected to an Arduino UNO, which sends data via a serial connection. The app serves as the project interface, prompting users with questions at 10-second intervals, recording serial data for analysis, and saving answers in separate CSV files.

Pulse Oximeter

A specially designed and 3D-printed piece holds a finger using an elastic sleeve, providing a dark backdrop for the LED lights to measure resistance accurately. This setup ensures reliable sensor readings.

Data Analysis

Collection

• Data collection is done through Arduino via a serial connection using PySide6 for handling data collection.

Analysis

• Used Python3 with libraries like SciPy, Matplotlib, Pandas, NumPy, and PyKalman for scientific computing.
• Processed the raw GSR and PPG signals to obtain 9-second signals from each question interval, resulting in 90 samples per recording for training.

Example Data

The CSV data obtained after processing looks like this:

Timestamps,GSR_Data,Red_PPG_Data,IR_PPG_Data,Processed_PPG_Data,SpO2,BPM
2.0,0.916,249217,195520,-81.629503,103.36753940548783,78.94736
2.075,0.937,249303,195525,-117.22019,101.36651055170812,78.94736

Training and Dataset

Dataset

• Consists of around 300 questions from 4 difficulty categories.

Training

• Initial training with 91 recordings showed an accuracy of about 73.68%.

Issues and Improvements

• The model's accuracy could be improved by collecting more data and refining the training dataset.
• Occasionally, the pulse oximeter sensor stops working due to cable issues, requiring a reconnection of the Arduino.

More images

Hardware

Software

3D model used for printing the oxidation sensor case