Core Concepts
The authors propose a novel hyperbolic one-class classification framework for face anti-spoofing that outperforms state-of-the-art methods on multiple benchmark datasets.
Abstract
The paper addresses the problem of face anti-spoofing (FAS), which aims to detect spoofing attacks on facial recognition systems. The authors argue that FAS should be treated as a one-class classification task, where the model is trained only on real samples and needs to detect unknown spoof attacks during inference.
The key contributions are:
Formulating FAS as a one-class classification problem and proposing a novel hyperbolic one-class classification framework called Hyp-OC.
Introducing two novel loss functions, Hyp-PC (Hyperbolic Pairwise Confusion loss) and Hyp-CE (Hyperbolic Cross Entropy loss), that operate in the hyperbolic space to learn discriminative features for one-class FAS.
Employing Euclidean feature clipping and gradient clipping to stabilize the training in the hyperbolic space.
Extensive experiments on five benchmark datasets demonstrating that Hyp-OC significantly outperforms state-of-the-art one-class FAS methods.
The authors first extract facial features using a VGG-16 network pre-trained on VGGFace. They then use a fully connected neural network for dimensionality reduction and map the features to the Poincaré Ball, a hyperbolic space. The Hyp-OC classifier, which operates in the hyperbolic space, is trained using the proposed Hyp-PC and Hyp-CE losses.
The results show that Hyp-OC achieves substantial improvements over previous one-class FAS methods, with an average HTER (Half Total Error Rate) reduction of 7.493 on intra-domain testing, 4.231 on leave-one-out inter-domain testing, and 2.778 on single-source-single-target inter-domain testing. The authors also conduct ablation studies to analyze the impact of different components of the proposed pipeline.
Stats
The mean of the Gaussian distribution used to sample pseudo-negative points is adaptively updated using a running mean of the real samples' features.
Euclidean feature clipping is used to address the vanishing gradient problem in the hyperbolic space.
Gradient clipping is employed to regularize the parameter updates in the hyperbolic space.
Quotes
"In the real world face anti-spoofing scenario, spoof samples are infinitely variable, which makes the task of FAS inherently complex."
"Hyperbolic space with a negative curvature allows for the learning of discriminative features owing to the nature of exponential growth in volume with respect to its radius."