Research on the recognition of USB electromagnetic features based on self-supervised loss control

The difference in the production of electromagnetic equipment will make it have unique electromagnetic characteristics, even if the same type of equipment, due to the environment in which it is located, the use of time or intensity of the different, will also lead to a certain degree of variability in the electromagnetic characteristics of its generation. This paper examines these subtle and unique differences by utilizing a self-supervised model to identify the individual electromagnetic characteristics of USB devices. Furthermore, the distinctive characteristics of electromagnetic radiation signals emitted by USB devices are extracted through a series of comparative experiments. In the experimental process, it was observed that: firstly, When the number of points in a single training data set is doubled, model recognition performance can be improved by about 2%; secondly, the larger the amount of data used, the greater the optimization of model performance achieved through the use of model training data. Nevertheless, the differentiation of noise is less pronounced, while the characteristics of heartbeat packets are more obvious and distinct. Consequently, the data files containing heartbeat packets demonstrate a more pronounced improvement in model performance. Thirdly, the model performance can be further enhanced by selecting the model loss threshold, provided that an adaptive threshold is obtained. Finally, the self-supervised method selected in this paper does not necessitate manual labeling, and self-labels data through clustering, thereby reducing the reliance on human and material resources. Furthermore, The self-supervised method selected in this paper does not necessitate the manual labeling of information, instead automatically classifying the data through clustering. This approach offers significant cost savings in terms of manpower and material resources, while also demonstrating high efficacy and a high recognition rate. Furthermore, the features extracted through self-supervision exhibit enhanced robustness.