Abstract: In order to solve the problem of feature randomness and feature drift in depth clustering model, a depth automatic encoder embedding clustering algorithm based on single network optimization is proposed. The strong competition relationship outside a single network was obtained through the discriminator, so as to avoid the random cost of features and feature drift. An adversarial training was further introduced to realize the tradeoff between feature randomness and feature drift, and two new indicators were introduced to evaluate the level of feature randomness and feature drift respectively. Experiments on several datasets prove the superiority of the proposed method in clustering applications.