Few-shot classification categorizes objects with minimal training data, making it valuable when large datasets are impractical. Models are trained on a base set with many samples per class and tested on a novel set, where they classify new samples using only a few examples per class. Since base and novel classes are distinct, models must generalize to unseen classes while training on the base set, making meta-learning more effective than traditional classification methods. State-of-the-art techniques improve generalization by pretraining on large datasets, followed by applying meta-learning to further enhance performance. However, we argue that although meta-learning is effective for few-shot tasks, models often overfit to the base classes, reducing performance on novel classes, even with pretraining. To address this issue, we propose two techniques in the meta-learning phase to reduce overfitting and improve generalization. First, we mask important parts of the sample to prevent the model from over-relying on specific features. Masking is applied using attention scores in ViT-like backbones or class activation maps in CNN-based backbones. Using the masked samples, we apply contrastive loss to prototypical network training, reducing the distance between a sample and its class prototype while increasing the distance to prototypes of other classes. The proposed method is applicable regardless of the backbone, whether a pretrained model is used, or whether the approach is inductive or transductive. We conduct experiments on various benchmark datasets and configurations to demonstrate the effectiveness of our method.