In view of an improvement of the recognition accuracy of deep learning models in the field of camera source recognition, an adaptive attention dense network structure has thus been designed, with an adaptive weighted attention method proposed. The designed network structure includes four parts: prepossessing structure, dense connection structure, attention structure, and regularization structure. The adaptive weight attention method is adopted in the attention mechanism structure so as to introduce adaptive weight factors for an parameter adaptive optimization, thus obtaining the optimal attention weight that adapts to different data characteristics and improve the model's feature learning and feature expression capabilities. Comparative experiments and ablation experiments are conducted on two classic datasets. In the comparative experiment, a comparison is made between the designed network and three classical networks, with the experimental results showing that the recognition performance of the designed network is improved by at least 5.547% and 9.283% on two datasets, respectively, compared to the other three networks. The results of the ablation experiment show that the recognition performance of the proposed method is at least 1.143% and 5.000% higher than other ablation methods on two datasets, respectively.