Determining the Optimal Number of User Access Modifications in ECG-Based Authentication Systems Using a Transfer Learning Approach

A key challenge in processing electrocardiogram (ECG) signals for personalized biometric authentication systems is the efficient retraining of deep learning models when new user data become available. This study employs a transfer learning approach combined with the VGG۱۶ architecture to investigate the permissible threshold for repeatedly integrating data from a single user—assigned with varying access labels—into a pre-trained base model. The base model was initially trained on ۴۶ users from the MIT-BIH database. For each target user, data were simulated across ۱۰۰ iterations with random label assignments and incrementally incorporated via gradual fine-tuning. The results reveal that the final model's performance improves with the addition of up to approximately ۴۴ to ۴۶ target models—a quantity comparable to the number of initial users—after which performance declines. The optimal range for adding new data from a single user was identified as ۱۱ to ۱۶ iterations. Within this interval, the model achieves an accuracy exceeding ۹۹% without exhibiting catastrophic forgetting. These findings confirm that while transfer learning offers an efficient mechanism for frequent updates in ECG-based biometric systems, it possesses a finite capacity for accommodating repeated modifications. Determining this threshold provides valuable practical guidance for the development of security and health monitoring systems that must balance functional flexibility with overall model stability.

کلیدواژه ها:

Transfer learning ، Electrocardiogram (ECG) Signal ، Catastrophic Forgetting ، Biometric Authentication