Elevating Wearable Sensor Authentication with Hybrid Deep Learning and Squeeze-and-Excitation

Smartphones often contain users’ private and sensitive information. Thus, continuous verification of identity is crucial to prevent unauthorized access. While sensor-based solutions have been explored before, recent advancements in machine learning have provided alternative deep-learning approaches for constantly checking if the smartphone’s user is genuine. We put forward a hybrid deep convolutional neural network model named SE-DeepConvNet, which includes squeeze-and-excitation elements to improve representation learning from streams of smartphone sensor information. Using a benchmark human movement dataset, our experiments show state-of-the-art performance boosts with SE-DeepConvNet, realizing 99.78% accuracy and 0.38% equal error rate for user verification – surpassing other evaluated deep learning models. Our distinctive combination of the squeeze-and-excitation system and deep convolutional neural network framework enables optimized feature extraction from sequence information for accurate constant verification.