Comparing the Performance of MobileNet and AlexNet Deep Learning-based Models for Facial Recognition

Abstract

Increasing sophistication in deepfake manipulations, identity fraud and unauthorized access underscores the need to transition security systems to intelligent and adaptive models. Research has brought deep learning to the forefront of architectures capable of learning highly discriminative representations for facial recognition and identity verification. This study evaluates the performance of two pre-trained convolutional neural networks, MobileNet and AlexNet with the use of Transfer Learning, applied to the recognition of authorized and unauthorized individuals from a customized dataset of 1000 images obtained from 20 individuals. Multi-task Cascaded Convolutional Networks (MTCNN) were employed for face and landmark detection. TensorFlow, Keras, PyTorch deep learning frameworks and OpenCV, Python Imaging Library (PIL), and scikit-image for Image Processing written in Python were used, while the performance analysis involving training (70%), validation (k-fold, k=10) and testing (30%) using confusion matrix, accuracy, precision, recall and F1-score performance metrics was done on the different images created from original dataset. Evaluation results reveal distinct trade-offs between the two architectures with confusion matrices results for AlexNet achieving better per-class accuracy in certain categories (up to 99% and 97% for authorized individuals) and demonstrated stronger confidence in rejecting unauthorized users (95% and 91%). In the performance metrics results, MobileNet exhibited higher overall accuracy (83%), precision (0.85), recall (0.83), and F1-score (0.83), with better generalization across classes and fewer false positives. These findings indicate that while AlexNet may be preferable in high-resource environments prioritizing maximum recognition rates, MobileNet offers a more balanced trade-off between accuracy, computational efficiency and generalization, making it more suitable for real-world deployment where efficiency and reliability are critical. In conclusion, this comparative analysis provides insights into model behavior under customized datasets, with implications for extending similar architectures toward multimodal biometric authentication and deepfake detection involving facial, voice, and image manipulations. The findings also highlight the necessity of aligning model selection with application requirements, ensuring robust and scalable solutions in evolving security contexts.