Rehearsal-Free and Efficient Continual Learning for Cross-Domain Face Anti-Spoofing | JoVE Visualize

Area of Science:

Computer Vision
Artificial Intelligence
Machine Learning

Background:

Face Anti-Spoofing (FAS) models face evolving spoofing techniques, necessitating robust continual learning strategies.
Existing methods for mitigating Catastrophic Forgetting (CF) in FAS are inadequate for subtle spoofing cues and lack generalization enhancement.
Catastrophic forgetting in FAS is linked to inconsistent attention patterns within Vision Transformer (ViT) models.

Purpose of the Study:

To develop an effective FAS continual learning strategy that mitigates Catastrophic Forgetting (CF) and improves generalization to unseen spoofing attacks.
To address the limitations of current CF strategies in handling the imperceptible nature of spoofing information.
To propose a method that extracts generalized spoofing texture information and maintains attention consistency.

Main Methods:

Proposed an instance-wise dynamic central difference convolutional adapter module with weighted ensemble for Vision Transformer (ViT) fine-tuning.
Introduced attention consistency regularization to alleviate Catastrophic Forgetting (CF) by learning and reusing ViT attention matrices.
Developed new protocols and conducted extensive experiments for validation.

Main Results:

The proposed method demonstrates superior performance in alleviating Catastrophic Forgetting (CF) during continual learning for Face Anti-Spoofing (FAS).
The approach significantly enhances the generalization capability of FAS models to unseen spoofing domains and attack types.
The method effectively extracts generalized spoofing texture information and maintains attention consistency across learning sessions.

Conclusions:

The developed dynamic adapter and attention consistency regularization are effective in combating Catastrophic Forgetting (CF) in Face Anti-Spoofing (FAS).
The proposed strategy offers a promising solution for building robust and adaptable FAS systems capable of handling evolving threats.
The findings highlight the importance of attention mechanisms and texture information for generalized spoofing detection in continual learning settings.