Abstract: In addressing the issue of counterfeit automotive components, traditional anti-counterfeiting technologies struggle to meet high-precision identification requirements due to their susceptibility to replication. This paper proposes an electromagnetic fingerprint recognition method based on a physical-constrained multimodal feature fusion network. By designing a multi-branch network architecture, the method processes power-frequency, differential, spatial, and high-order statistical features, while incorporating constraints derived from electromagnetic physical laws to achieve accurate anti-counterfeit identification of automotive components. Experiments employ Electronic Product Code (EPC) passive ultra-high frequency (UHF) tags and utilize a radar reader to construct a multidimensional matrix model of radio frequency signals. Signals reflected from the tags and transmitted through automotive components are collected within the 840–925 MHz frequency band, followed by electromagnetic fingerprint data extraction and analysis. Results demonstrate that the proposed method achieves an accuracy of 97.4% and a recall rate of 95.7% in classifying genuine and counterfeit automotive components, outperforming traditional Support Vector Machine (SVM) by 6.5%. This study provides a highly robust solution for automotive component anti-counterfeiting, offering significant practical engineering value.