The rapid development of deep forgery technology has improved the quality of generated pictures and videos to mirror reality. However, it has brought huge security risks to society. In view of the large parameters used in existing detection methods, deep network, complex model structure, etc., this paper first optimizes the classic detection model XceptionNet in the forensics field and proposes a lightweight forensic model Xcep_Block8 that reduces the model parameters while maintaining high detection accuracy. Second, we improve the solution of the unevenness of positive and negative samples by increasing the sampling probability of samples with fewer categories to solve the problem of unbalanced categories. Finally, we employ the hybrid data enhancement method, CutMix, to improve the linear expression between samples. The experimental results show that the test results of the proposed model are about 1.01 percentage points higher than the baseline results. Additionally, it has certain advantages compared with other methods in terms of parameter quantity.