back to work
Case studyB.Tech Thesis · IIIT SriCity
FedFV-CV
Federated Deep Learning for Biometric Auth
1.21%
Equal Error Rate
122.6K
images · 5 clients
> FedAvg
on EER benchmark
The problem
Finger-vein biometrics are sensitive — centralising raw vein images to train a model is a privacy risk. The goal: train an accurate authenticator without the raw data ever leaving each client.
Approach
01
Privacy-preserving federated training
MobileNetV2 trained across 5 clients over 122,600 images. Only model updates are shared; raw vein images never leave the device.
02
Custom FedWPR aggregation
A weighted-performance aggregation scheme that outperformed the FedAvg baseline on non-IID client splits, where naive averaging degrades.
Stack
PyTorchMobileNetV2Federated Learning
Next case study
slackAgent
AI-Powered Slack Bot with RAG