04
Liveness — Face anti-spoofing
The sequel to Deep Scan: telling a real face from an attack (printed photo, screen). Fine-tuning an ImageNet-pretrained ResNet18 for spoof / non-spoof classification, with a Streamlit interface for testing.
Stack
What happened?
- 01
The question
Deep Scan recognises a face. But what happens when you show it a photo of that face? Can a model learn to tell a real face from an imitation?
- 02
The problem
To a recognition system, a printed photo or a face on a screen looks a lot like the real person. Without spoof detection, a simple photo can be enough to fool the system.
- 03
The hypothesis
An attack leaves traces the eye barely notices: paper texture, screen glare and pixels. A pretrained ResNet fine-tuned on spoof / non-spoof examples should learn to spot them.
- 04
The build
Fine-tuning an ImageNet-pretrained ResNet18 (with an option to freeze the backbone), an inference script, unit tests, and a Streamlit interface showing the predicted class and confidence score.
- 05
The next frontier
Reach reliable results, measure robustness against different attack types, then plug the model in front of Deep Scan.