Electromagnetic (EM) emissions from electronic devices can inadvertently carry sensitive information, posing significant threats to information security. EM fingerprinting techniques have become vital for security inspection and leakage source localization, yet existing approaches often suffer from poor adaptability across sampling rates and insufficient extraction of high-frequency features.

This study aims to develop a robust EM fingerprint recognition method that maintains high accuracy across different sampling rates while effectively capturing high-frequency characteristics, thereby improving security detection and adaptability in practical scenarios.

We propose an enhanced neural network architecture, termed ELEC-TDNN, which integrates a channel attention mechanism with multi-scale temporal modeling capabilities. A local signal enhancement layer is introduced to improve the representation of subtle EM features. Experiments were conducted on a self-constructed dual-sampling-rate USB device EM emission dataset (1.25 GHz and 500 MHz) to evaluate performance. The evaluation used equal error rate (EER) as the primary metric to measure recognition accuracy under varying frequency conditions.

The proposed ELEC-TDNN achieved superior adaptability and accuracy compared with conventional methods. At 500 MHz, the model attained a minimum EER of 0.35%, while in the high-frequency 1.25 GHz scenario, it achieved an EER of 5.23%. These results indicate that the architecture effectively preserves recognition performance despite significant differences in sampling rates.

By combining attention-based channel feature selection, multi-scale temporal modeling, and local signal enhancement, the method addresses both cross-sampling-rate adaptability and high-frequency feature extraction challenges. This work demonstrates practical value in enhancing EM security detection systems and offers a scalable approach for future EM analysis in multi-rate environments.