Recent deep learning-based steganalysis models have shown remarkable success in detecting steganographic algorithms matching their training data, yet struggle with previously unseen algorithms. This performance degradation primarily stems from the loose feature distributions and ambiguous class boundaries learned by current approaches. To overcome this limitation, we introduce a novel framework incorporating two key components: a feature clustering module designed to increase intra-class compactness and a distance constraint module that effectively separates inter-class feature centers. These components work synergistically to optimize feature distributions and sharpen decision boundaries for improved steganalysis performance. Our proposed FCDC framework integrates these modules while maintaining computational efficiency, adding negligible overhead to existing models. Extensive experimental validation across standard datasets including BOSSbase&BOWS2 and ALASKA#2 demonstrates consistent generalization improvements when applied to SRNet, SiaStegNet and LWENet architectures. The framework achieves performance gains up to 7.4% points without significant increases in computational cost or memory requirements, representing a meaningful advance in practical steganalysis applications.