Coursework: Machine learning research
Federated face-recognition robustness
A team study of how adversarial attacks degrade smart-home face recognition trained with federated learning, and a hybrid defense that combines denoising autoencoders with a reverse-attack step.
What it is
Face recognition in a smart home can be fooled by small, deliberate changes to an image. When the model is trained across devices with federated learning, the attack surface and the defenses both change.
The project started from the FLATS enhancement setting. We tested attacks (FGSM, FFGSM and Square), and explored denoising autoencoders and a reverse-attack strategy as defenses.
My part
Team lead. I led a team of four and coordinated experiment design, result analysis and the final presentation.
Graduate course project done as a team of four, which I led.
Stack
- Python
- PyTorch
- Federated learning
- Denoising autoencoders