Social recommendation systems utilize social relationships to improve recommendation performance. However, most existing methods primarily focus on pairwise user connections, often neglecting the dynamic, high-order, and collective influence exerted by social groups. In this paper, we propose a novel Group Information Enhanced Social Recommendation (GESR) model that explicitly captures high-order interactions between users and social groups to boost recommendation quality. Specifically, we first introduce a social network refinement mechanism to denoise the raw social networks by filtering out weak or irrelevant social relationships. Then, we design a bidirectional user-group attention mechanism to model the mutual influence between users and their affiliated social groups, capturing both top-down and bottom-up interactions. Furthermore, we incorporate a multi-layer graph attention network-based high-order information fusion strategy to aggregate multi-hop relation information, enabling the model to learn expressive user and item representations. Finally, extensive experiments on two public datasets demonstrate that GESR significantly outperforms SOTA baselines, validating the effectiveness and superiority in social recommendation.