Recently, machine unlearning has been essential for large ML models to reduce the computational complexity of retraining from the beginning. In literature, many researchers are working in a similar direction; it is still an unsolved problem because of the non-availability of a generalised framework. Exploration of a generalised framework for class-wise unlearning of data along with learning of the new class of data, which, unlike traditional models that add up noise for unlearning, hides the classes to be unlearnt. This framework consists of two sub-models viz., First-level and Second-level model. The First-level model is used to identify whether a given sample belongs to existing classes or newly added classes, while the Second-level model is used to classify the given sample with newly added classes. For generalization of the proposed framework, in our work we have explored six different domain standard image datasets with the presence of both conventional and convolutional models. Subsequently, we empirically prove that sub-model's performance directly affects the performance of the proposed framework. The experimental evidence recommends that the proposed framework is a generalized framework for class-wise unlearning and new learning. Out of the six considered datasets and their corresponding models the Face dataset with (FaceNet + SVM) model and MNIST dataset with (Binarisation + CNN) model outperformed the rest of the dataset models with lesser error rates being achieved in the First-level and Second-level sub-models of our proposed framework.