Bishwa, let me try to explain it differently and let me know if helps
Using GAN we would like to generate data which is a close replica of real data. Data can be text, images, videos, music etc.
GAN uses Generator and Discriminator to achieve this goal of generating replica of real data.
During training we train Discriminator on real data as well as the fake data generated by Generator. This will help Discriminator to get best at the job of classifying input data as real or fake.
Initially Discriminator will be good at identifying fake from real as Generator is still learning and Generator has not mastered the art of generating fake data yet.
As the classification error is constantly fed back to Generator. Generator will get better at generating data as close as the real data.
As we keep training the Generator and the Discriminator we can see that the Generator gets better at generating fake images that looks real.
Real data helps Discriminator learn the feature of the real data. This is then used by Discriminator to classify data as real or fake. Discriminator does this by calculating the loss function for real and fake data