Handwritten signature verification is crucial for security in financial and administrative processes but faces challenges due to significant intra-class variations and low inter-class distinctions, especially when dealing with skilled forgeries. To address these issues, this paper proposes DSA-Net, a novel selective attention module integrated into an enhanced SigNet architecture. The attention module constructs a selective attention matrix through batch processing and Top-k selection strategy, which effectively focuses on the local features of the signatures and enables the model to better discriminate between genuine and skilled forged signatures. Additionally, this paper proposes a novel dynamic contrastive loss function. This loss function dynamically optimizes the distribution of genuine and forged signatures in feature space, enabling the model to more effectively distinguish signatures of varying difficulty and enhance its discrimination performance. Experiments on datasets like CEDAR and BHSig260 show that our method achieves comparable or better accuracy than state-of-the-art approaches.