Edge AI / Privacy · 2022–2023
Federated Learning with Object Detection
University of Essex · Dissertation
01
Problem
Deploying vision models on IoT fleets raises privacy and bandwidth costs when raw images leave the device.
02
Challenge
Combine federated learning with computer vision so edge devices can improve models without centralizing sensitive imagery.
03
Solution
Researched and validated a distributed AI approach for IoT—federated training patterns paired with object detection for privacy-preserving edge intelligence.
04
Impact
Demonstrated a practical pattern for privacy-preserving vision on constrained devices; the direct research foundation for NxtPerson's on-device crop-grading roadmap and AUTOEMATIO's edge monitoring work.
Approach
- 01
Reviewed federated learning and edge-AI literature to identify architectures suited to bandwidth- and privacy-constrained IoT deployments.
- 02
Implemented a federated training loop where model updates, not raw images, were shared between simulated edge nodes.
- 03
Paired the federated setup with an object-detection model to test accuracy trade-offs against a centralized baseline.
- 04
Evaluated the approach against MSc modules in machine learning, neural networks, and robotics for methodological rigor.
Outcomes
- Working federated learning + object detection prototype for edge devices
- Documented accuracy trade-offs versus centralized training
- Reusable architectural pattern later applied to agritech and industrial edge scenarios
- Dissertation submitted as the capstone of the MSc Artificial Intelligence programme