XAI-FR: Explainable AI-Based Face Recognition Using Deep Neural Networks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Face recognition systems, while advanced, often lack transparency in their decision-making processes.
• Understanding the features driving face recognition is crucial for user trust and system validation.

Purpose of the Study:

• To evaluate the interpretability of deep neural network-based face recognizers.
• To demonstrate how explainable AI tools can clarify the feature importance in face recognition models.

Main Methods:

• Employed deep learning models: LeNet-5, AlexNet, Inception-V3, and VGG16.
• Utilized the Local Interpretable Model-Agnostic Explanations (LIME) technique for interpretability analysis.
• Tested on benchmark datasets: Yale, AT&T, and Labeled Faces in the Wild (LFW).

Main Results:

• Local Interpretable Model-Agnostic Explanations (LIME) successfully identified visually significant facial features.
• The interpretability method highlighted features crucial for the accurate recognition of individuals across different models.
• Demonstrated consistency in feature highlighting across various deep learning architectures.

Conclusions:

• Explainable AI, specifically LIME, can significantly enhance the transparency of deep learning face recognition systems.
• The findings provide a method for users to understand the basis of AI-driven facial identification.
• This research contributes to building more trustworthy and interpretable AI in biometrics.