Existing social denoising strategies often rely on user-item interaction data for preference-guided denoising, basically ignoring the problem that the interaction data itself may be noisy, and there are limitations in the hardness control of existing negative sampling strategies, which will reduce the model’s recommendation performance. To address these issues, an information bottleneck-based social recommendation optimization model, IB-SRO, was proposed. First, an information bottleneck-guided dual-view collaborative denoising module was introduced, in which social network noise and user-item interaction noise were asynchronously optimized through a multi-stage learning strategy. A collaborative optimization method guided by the information bottleneck was applied to sequentially optimize the two types of noise, increasing the independence between different views and enhancing the denoising effect. Moreover, a novel difference-aware negative sampling strategy was designed. By combining similarity and difference-aware functions, the strategy was enabled to adaptively select appropriately hard negative samples based on the predicted scores of positive samples, enhancing the learning of user preferences and thereby improving the model’s performance. Experimental results showed that compared with the state-of-the-art models, IB-SRO achieved average improvements of 6.87%, 6.72%, and 2.73% on the Epinions, Yelp, and Douban-Book datasets respectively in Top-N recommendation tasks, which verified the effectiveness of the model.