Federated Generalized Face Presentation Attack Detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Face presentation attack detection (fPAD) is crucial for secure face recognition.
• Data privacy concerns limit the sharing of real and spoof face images for training.
• Existing fPAD models struggle with generalization to diverse spoofing techniques and data distributions.

Purpose of the Study:

• To propose a federated face presentation attack detection (FedPAD) framework to address data privacy challenges.
• To develop a federated generalized face presentation attack detection (FedGPAD) framework for improved generalization to unseen attacks.
• To evaluate the effectiveness of FedPAD and FedGPAD through extensive experiments.

Main Methods:

• FedPAD: Local training of fPAD models at data centers and aggregation of model updates by a central server without accessing private data.
• FedGPAD: Introduces federated domain disentanglement, decomposing models into domain-invariant and domain-specific parts.
• Server aggregates only domain-invariant parts in FedGPAD to enhance generalization.

Main Results:

• FedPAD successfully enables collaborative fPAD model training while preserving data privacy.
• FedGPAD demonstrates superior generalization capabilities compared to standard federated learning approaches.
• Extensive experiments provide insights into federated learning strategies for fPAD.

Conclusions:

• Federated learning offers a viable solution for privacy-preserving fPAD.
• FedGPAD framework significantly improves the generalization of fPAD models to novel spoofing attacks.
• The proposed methods pave the way for more robust and secure face recognition systems.