The invention provides a password attack evaluation method based on conditional variational self-coding. The method comprises the following steps of constructing a variational self-coding model, constructing a conditional variational self-coding model, constructing a password attack model, and comprehensively utilizing the conditional variational self-coding model and the password attack model. According to the invention, a conditional variational self-coding model, and the condition characteristics in the personal information of the user, such as the user name, the mailbox address, the telephone number, etc. are used to train the cryptographic attack model, a bidirectional GRU recurrent neural network and a CNN text convolutional neural network are respectively used at the encoder end, sothat the code of a password sequence and personal information of a user and abstract extraction of characteristics can be achieved, the distribution of password data and a character combination rulecan be effectively fitted, the high-quality guessing password data can be generated, and the method has a remarkable effect on improving the password strength and safety of the user.