Abstract: An attention-guided multi-dimensional feature fusion identification method based on a distributed radar system is proposed in this paper. In this method, a multi-feature fusion network is designed to focuse on different echo characteristics, and multi-modal information is projected into a high-dimensional space for effective separation. The network comprehensively utilizes the energy consistency of target echoes, velocity projection consistency, spatial positioning correlation, and spectral distortion features formed by genuine reflections, paying attention to features of signals at different scales by means of multi-scale convolution; moreover, it enhances the representational capacity of network by establishing an attention mechanism among different scales and features. The identification performance is validated and compared by constructing a dataset based on electromagnetic simulations, and designing various signal to noise ratio and baseline length conditions. Experimental results demonstrate that under challenging conditions of a signal to noise ratio of -6 dB and a short baseline of 2 km, the proposed method achieves a recognition rate of 97.4 %, showing significant improvement over traditional methods. Component-wise and feature-wise ablation studies confirm the effectiveness of key designs, including the attention mechanism, multi-scale convolution, and multi-feature fusion. As the proposed method is grounded in physical characteristics without relying on specific signal modes, it offers a more robust solution for true-false target identification in complex electromagnetic environments, thereby enhancing radar capability against deception jamming.