Privacy-Preserving Efficient Deep Learning Model for Skin Lesion Classification

Making informed treatment decisions and preventing major skin disorders depend on early diagnosis of the types of skin lesions. While there are traditional methods available, they often lack accuracy and struggle to handle complex multidimensional data, particularly in situations where resources are scarce. This work presents a "Privacy-Preserving Efficient Deep Learning Model" for skin lesion classification that employs 3 custom CNN methods which are additionally enhanced with differential privacy to provide security to the sensitive medical images. The suggested approach makes use of convolutional neural networks (CNNs) and differential privacy strategies to protect private patient information while preserving classification performance. Three modified CNN architectures are used to assess the model’s efficacy using important measures such as accuracy, precision, recall, and F1-score. Notably, the model shows that data security measures do not impair predictive performance by achieving an accuracy of 97% with differential privacy included. For data augmentation, the study also uses conventional data augmentation processes like flipping, rotating, calling, cropping etc. which strengthens the model’s resistance to class imbalances. A time study validated the computational efficiency of the suggested method. This study demonstrates the feasibility of incorporating differential privacy into deep learning pipelines for medical imaging applications, while balancing accuracy and security.