Federated Self-Supervised Few-Shot Face Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Privacy-preserving machine learning is crucial for sensitive data like facial information.
• Federated learning (FL), self-supervised learning (SSL), and few-shot learning (FSL) are advanced techniques with potential for privacy-preserving face recognition.

Purpose of the Study:

• To develop and evaluate a systematic framework combining FL, SSL, and FSL for privacy-preserving face recognition.
• To quantify the performance trade-offs associated with privacy preservation in face recognition systems.

Main Methods:

• Self-supervised pre-training using SimCLR on CASIA-WebFace dataset in a federated setting.
• Federated few-shot fine-tuning on LFW dataset utilizing prototypical networks.
• Comprehensive evaluation across six architectures: ResNet, DenseNet, MobileViT, ViT-Small, CvT, and CoAtNet.

Main Results:

• The federated framework successfully preserves data privacy but incurs significant performance degradation (12-30% accuracy loss) compared to centralized methods.
• Convolutional Neural Networks (CNNs) demonstrated greater robustness under federated constraints than transformer-based architectures.
• Five-shot learning configurations offered an optimal balance between data efficiency and recognition performance.

Conclusions:

• Privacy preservation in federated face recognition comes at a substantial performance cost.
• CNNs are more suitable than transformers for federated few-shot face recognition tasks.
• Empirical benchmarks and insights are provided for deploying privacy-preserving face recognition systems, highlighting critical privacy-utility trade-offs.