Multi-data feature domain disturbance attack method based on GAN model

The invention provides a multi-data feature domain disturbance attack method based on a GAN model, and the method is characterized in that the GAN model is a model which comprises two neural networks, namely a generator and a discriminator, and achieves a mutual balance state through the mutual game between the generator and the discriminator. The multi-data feature domain means that the data existence form of the same object is not single, for example, radar data comprises a complex number format, a real number format and an RGB format, and the same audio frequency also has a time frequency domain format, a space frequency domain format and the like. According to the multi-data feature domain disturbance attack method based on the GAN model, a relative mean cross entropy loss function is provided, and the distinguishing condition of the same category is also considered under the condition of considering the distinguishing condition of a single sample. Meanwhile, through a cross-domain algorithm, perturbation