Numerous approaches for ego-vehicle 3D perception tasks have arisen as a result of the advancement of AI-assisted driving, but there has been little research on roadside perception. The roadside perspective is worthwhile to develop since it offers a global picture and a wider sensory range. While cameras offer semantic information, LiDAR offers exact 3-D spatial data. In 3D detection, these two modalities work well together. Nevertheless, since both the extraction and fusion process is not sufficiently accurate, additional camera data fails to enhance accuracy in certain tests. Kolmogorov-Arnold Networks (KANs), which are more appropriate for high-dimensional, complicated data, have recently been suggested as alternatives to MLPs. Kaninfradet3D, which improves the feature extraction along with fusion modules, is proposed in this study. KAN Layers were used to enhance the encoder and fuser modules of the model. Cross-attention was used to improve feature fusion, and visually comparisons showed that camera features had better merged. This solved the issue of unusually concentrated camera characteristics, which had a detrimental influence on fusion. Our method surpasses the benchmark by +1.40 mAP in the infrastructure portion of the TUMTraf V2X Cooperative Perception Dataset and by +9.87 mAP and +10.64 mAP in both perspectives of the TUMTraf Intersection Dataset. The results highlight the potential of employing KANs in roadside perception challenges by demonstrating that Kaninfradet3D can successfully fuse features.