Deep Tree Network for Object Classification

For the final project in UVic's CSC 486B: Introduction to Deep Learning for Computer Vision, my group implemented and modified a state of the art machine learning model.

A face spoof attack is an attempt to deceive a facial recognition system by using a fake face to either impersonate a genuine user or to hide an attacker's identity. Some common such attacks include the use of makeup, masks, or printed photographs to try to bypass recognition systems. Face anti-spoofing is used as a preliminary step to filter out these spoofs before they're sent to the recognition system. The paper we reproduced used a novel Deep Tree Network to try to identify face spoofs using techniques it was not trained against.

Contributions

Our primary contribution was to apply this model to a different binary classification problem. Instead of deciding whether an image was a genuine face or a spoofed one, we modified the model to decide whether an image was human-made (like a car, a toaster, or a house) or natural (like a tree, a lake, or an animal).

Our implementation also differed from the original paper's in a number of other ways:

• We ported the model from TensorFlow2 to PyTorch
• Instead of hardcoding the tree depth to 3, we made it a variable hyperparameter
• We removed the face anti-spoofing specific binary masks used in one of the primary network modules of the original model

Report

A more complete write-up including our methodology and results can be downloaded here.