A Wi-Fi CSI Based Low-Cost Security System with Self-Supervised Training for Open-Set Intruder Alerting

Overview

The goal of this project was to determine if it is possible to create a security system that uses changes in WiFi Channel State Information, collected from ESP32 sensors, to distinguish household residents from previously unseen intruders without relying on cameras or expensive hardware. A deep learning model (SE-ResNet10) was trained on Wi-Fi CSI data from the XRF55 dataset using a few-shot learning framework. Raw CSI signals were preprocessed with a Butterworth low-pass filter before being fed into the model. The pretrained model was then fine-tuned on a small set of household resident samples to adapt it to a specific deployment environment. To test its open-set intruder detection capabilities, incoming CSI windows were compared against resident profiles using a similarity score, and a threshold was applied to distinguish known residents from completely unseen intruders. The threshold was swept across its full range to identify the optimal operating point balancing false acceptance and false rejection rates. The system achieved strong separation between residents and intruders despite having zero exposure to intruder data. The system successfully distinguished between residents and intruders. At the optimal threshold of 79%, it achieved 92.66% accuracy, with a False Accept Rate of 4.4% and a False Reject Rate of 10.4%. Furthermore, the concept of continuous fine-tuning puts the model on a path to achieve 99%+ accuracy. My results show that a model can achieve open-set intruder detection with high accuracy, suggesting that existing home Wi-Fi infrastructure could provide a low-cost and privacy-preserving home security solution.

Competition history

• CSEF 2026 Computational Science (Senior Division) · Entry S-07-21

Related projects

Closest projects by meaning, across every fair and year in the corpus.

Browse more like this

• All Computational Sciences projects
• CSEF Computational Sciences projects
• Every CSEF 2026 project