Abstract: To achieve high-quality mapping between surface 2D gravity anomaly observations and subsurface three-dimensional (3D) density geological bodies, this paper proposes a novel deep learning network architecture—Fusion Kernel Attention Network (FKANET). Compared to conventional neural network models previously used for 3D gravity inversion, FKANET can improve the inversion accuracy of both geometry and density amplitude of subsurface 3D density geological bodies while significantly reducing the number of network parameters. Furthermore, to address the widespread noise in actual gravity anomaly observations, this study reconstructs an improved network model based on a multi-task framework using the FKANET architecture, aiming to achieve simultaneous inversion and denoising of noisy data. Both tasks share the same encoder to extract gravity anomaly features from the raw gravity anomaly observations. In contrast, two independent decoders are used to perform denoising and 3D geological body inversion individually. Besides the adaptive denoising capability to effectively suppress the noise interferences in gravity anomaly observations, this multi-task framework poses sufficient constraints on the inversion process to enhance the model's ability to resolve subsurface 3D density geological bodies. These characteristics collectively enhance the stability and accuracy of inversion results. This method was applied to the inversion of field data from the Sannicolas region in central Mexico, demonstrating that the multi-task FKANET can achieve high-precision, real-time inversion of geological boundaries and density while also exhibiting excellent noise immunity.