Waves Don’t Lie: Leveraging Test-Time Training and Kolmogorov Arnold Networks for EEG-Based Biometrics

There is an increasing interest in the application of EEG signals for user identification within healthcare and Internet of Things (IoT) systems, driven by the need for more secure and reliable biometric authentication methods. One of the challenges in EEG-based identification is the variability in signals due to cognitive states and environmental factors. This study explores the effectiveness of three deep learning models—Kolmogorov Arnold Network (KAN), Test Time Training (TTT)-enhanced models, and Multi-Layer Perceptron (MLP)—for EEG-based biometric authentication. KAN helps in reducing overfitting while, TTT improves the robustness and generalization of EEGbased identification by dynamically adjusting to variations in EEG signals. A publicly available EEG dataset augmented with newly collected data was used in this review to ensure comprehensive and robust model evaluation. The results showed that the TTT-based model outperformed the others, achieving an accuracy of $95.4 \%$, compared to $91 \%$ for CNN-LSTM and 92% for CNN-LSTM-KAN, across various trials. These findings demonstrate the potential of TTT as a powerful framework for practical and secure EEG-based biometric authentication.