Background: GANs are well-known for their ability to generate realistic fake sample data, which can be audio, images, and videos. The application areas of GANs have increased their popularity in recent years. The first and best feature of GANs is their learning nature, characterized by powerful learning. As GANs have a strong discriminating ability to differentiate fake data from real data. This thesis tries to use that discriminating ability in classifying tasks such as handwritten digit classification.

Objective: First, a literature review was conducted to identify the appropriate performance metrics for comparing the classifiers. To train different GANs and to compare the performance of each GAN as feature extractors for handwritten digits classification with traditional algorithms such as SVM, Random forest and CNN.

Methods: We performed a literature review to determine metrics to compare the performance of the classifiers and understand which traditional algorithms are mostly used in Handwritten digit classification task. Experiment is conducted using DIDA handwritten digits data set with DCGAN, ACGAN, and CGAN algorithms.

Results: The result of the literature review indicates accuracy, precision, and recall metrics can be used to compare classification algorithms. The results of the experiments conclude that ACGAN, with a classification accuracy of 76.6\%, outperforms CGAN and DCGAN-based classifiers with an accuracy of 69.6\% and 74.9\%, respectively. The results of SVM, Random forest and CNN are 82\%, 78.3\% and 94\% respectively.

Conclusions: After analyzing all the results, we concluded that CNN outperforms GAN-based methods. However, this thesis concludes that the GANs can also be used as decent feature extractors in classification tasks as the performance of the GAN-based classifiers cant compete with the machine learning classifier such as SVM, Random Forest and CNN on DIDA dataset.