Skilled forgery signatures are very similar to genuine signatures, which seriously reduces the verification accuracy. Therefore, based on DenseNet, which has the advantages feature reuse, this paper proposes a multi-scale attentional feature fusion and cross-writer network to improve the performance of writer independent offline handwritten signature verification. Firstly, attentional feature fusion module is introduced to fuse multi-scale features to enhance feature representation capability. Secondly, cross-writer attention is used to highlight key areas of multi-scale feature differences between reference signatures and query signatures. In addition, in order to compensate for the shortcomings of convolutional neural networks in global information extraction, recurrent criss-cross attention is introduced to obtain contextual dependencies between pixels in signature images. Finally, a joint loss function is used to overcome the problem of imbalanced positive and negative signature pairs and improve the verification accuracy. The verification experiment was conducted on publicly available handwritten signature datasets Bengali and Hindi, and the proposed method has significant advantages compared to state-of-the-art results.