TransFace++: Rethinking the Face Recognition Paradigm With a Focus on Accuracy, Efficiency, and Security.

Area of Science:

• Computer Science, Artificial Intelligence, Machine Learning

Background:

• Current deep learning face recognition (FR) models, typically using Convolutional Neural Networks (CNNs) with RGB images, face challenges in capturing global facial features, efficiency, and privacy.
• CNNs struggle with global feature representation and local feature correlations, while RGB inputs increase computational costs and pose privacy risks if compromised.

Purpose of the Study:

• To propose novel FR frameworks, TransFace and TransFace++, that overcome the limitations of existing CNN-based models.
• To explore the application of Vision Transformers (ViTs) and raw image bytes for improved FR efficiency, security, and precision.

Main Methods:

• TransFace utilizes ViTs with novel data augmentation (Dominant Patch Amplitude Perturbation - DPAP) and hard sample mining (Entropy-guided Hard Sample Mining - EHSM) strategies tailored for facial structural information.
• TransFace++ processes raw image bytes directly, incorporating Topology-based Image Bytes Compression (TIBC) for feature extraction and Structure Information-guided Cross-Attention (SICA) to enhance geometric information perception.

Main Results:

• TransFace demonstrates improved performance by addressing ViT's limitations in large-scale FR datasets through specialized augmentation and mining techniques.
• TransFace++ achieves enhanced inference efficiency and user privacy by operating on image bytes, effectively mitigating information loss and improving generalization.

Conclusions:

• The proposed TransFace and TransFace++ frameworks offer significant advancements in face recognition technology.
• These novel approaches provide more efficient, secure, and precise FR solutions compared to traditional CNN-based methods, particularly for large datasets and privacy-sensitive applications.