<p>Breast cancer remains the most prevalent malignancy among women worldwide, where early and accurate screening is paramount for improving diagnosis rates and ultimately reducing mortality. Mammography serves as the cornerstone modality for breast cancer screening, offering essential diagnostic information through standardized multi-view imaging, typically comprising the craniocaudal and mediolateral oblique views for each breast. Nevertheless, current state-of-the-art deep learning approaches predominantly focus on a dual-view analysis restricted to a single breast, thereby failing to fully leverage the rich contextual information embedded across multiple ipsilateral views and consistently ignoring the potential correlations and symmetries between bilateral breasts, which can be critical for identifying asymmetries indicative of malignancy. To overcome these limitations, this paper proposes a cross-difference-driven dual-stream contrast multi-view network for mammogram classification (CDCM-Net). The network employs an innovative dual-module co-design architecture. It comprises cross-attention fusion module that effectively aligns and integrates features from ipsilateral views to capture their intrinsic consistency and complementarity, along with difference learning module that explicitly enhances discriminative feature representation by conducting structured comparisons between contralateral breasts to highlight suspicious divergences. Experimental results from rigorous evaluations on both the public DDSM and In-house datasets consistently reveal that CDCM-Net surpasses numerous existing advanced methods in breast cancer classification tasks. CDCM-Net provides an effective tool for the accurate screening of early-stage breast cancer.</p>