Skin cancer detection through dermoscopy images has remained challenging due to higher inter-class uniformity and intra-class diversity. Deep convolutional neural networks (CNNs) have recently obtained remarkable attention for automated skin cancer classification; however, most of these methods extract features from the global image of high resolution. One of the major drawbacks of these methods is the downscaling of input images to a low resolution, which leads to information loss. Moreover, the lack of a vast number of dermoscopy images is a concern. To overcome these issues and improve the classification accuracy, we propose a novel method using patch-based local deep feature extraction. In the proposed method, the features are extracted from different patches of a dermoscopy image using a pre-trained CNN model and are fused to preserve fine details. The kernel principal component analysis is then employed to select the prominent features, and a feed-forward neural network is finally used to detect skin cancer. The approach is validated using two benchmark datasets: ISIC 2016 and ISIC 2017. The experimental results show that the suggested method achieves promising results in comparison with state-of-the-art approaches. Moreover, a comparative analysis is made among contemporary pre-trained CNN models.