Molecular communication (MC) is an emerging paradigm for exchanging information via chemical signals. It holds great promise for building nanoscale networks within biological organisms. However, the presence of counting noise originating from the molecular diffusion mechanism severely limits the signal detection performance. Currently, research on reducing such signal-dependent noise (SDN) in MC via diffusion (MCvD) remains scarcely explored. Besides, the lack of channel models in MCvD hinders the effects of model-based denoising approaches. Against this background, we resort to the data-driven machine learning (ML) methods that fit MCvD systems for noise mitigation. Additionally, the simulated numerical results show that such an ML-based scheme outperforms its classical model-based counterparts.