Enterprise AI Analysis

This research introduces a privacy-preserving federated transfer learning framework for highly accurate skin disease diagnosis. It outlines a resource-efficient FL approach for recognizing and classifying skin illnesses. IoT-enabled devices collect and store skin disease images locally. The framework uses four strategies: (a) transfer learning with DNN classification, (b) feature extraction with pre-trained models and DNN classification, (c) federated learning with MobileNetV2 and DNN on IID and non-IID datasets, and (d) federated transfer learning with UNet-based feature extraction and DNN classification. Pre-processing involves resizing, grayscale conversion, and sharpening. The FL model safeguards data privacy by arranging statistics and secrecy across detached plans, making it suitable for submissions where secrecy is important. Model training uses transfer knowledge to apply information from previously trained models to new, related situations, speeding up training and reducing computational costs. Pre-trained architectures like DenseNet, VGG19, Xception, and UNet are used for feature extraction, while a Dense Neural Network (DNN) performs complex classification tasks by learning hierarchical features and disease-specific patterns. The study validates performance on both IID and non-IID datasets, with UNet+FL+DNN achieving the highest accuracy and resource efficiency.

This segment presents the results and comparative examination of the models used to identify skin illnesses, measured by accuracy, precision, recall, and loss. The proposed FL model was implemented on Google Colab using Python, with model training and testing on a GPU server. The HAM10000 dataset, comprising 11,253 dermatoscope images across seven skin infection types, was used for experiments. Hyperparameter tuning was optimized through empirical tuning and grid search, with a learning rate of 0.001, batch size of 32, and 100 epochs proving most effective. The study evaluated four models: transfer learning models (VGG16, Xception, EfficientNetB3, MobileNetV2), feature extraction models (DenseNet, VGG19, Xception, UNet with DNN classification), federated transfer learning models (MobileNetV2 + FL + DNN), and UNet + FL + DNN models. MobileNetV2 showed strong performance in initial transfer learning with 98.064% testing accuracy. UNet + FL + DNN achieved the highest accuracy across IID (99.528%) and non-IID (99.689%) datasets, along with the lowest loss. Resource utilization metrics (GPU memory, GPU process, CPU process, and virtual memory) indicated that FL-based models significantly reduce computational demands compared to traditional methods, with UNet + FL + DNN being the most resource-efficient while maintaining superior diagnostic accuracy.

The discussion highlights the benefits of federated transfer learning in skin disease diagnosis, particularly in terms of accuracy, resource efficiency, and data privacy. Integrating FL with DNN classification significantly improved performance across all models, with UNet + FL + DNN achieving the highest accuracy (99.689% for Non-IID data) and lowest loss (0.415% for IID data). This strategy outperformed MobileNetV2 and standalone UNet. The resource consumption analysis showed that FL-based models, especially UNet + FL + DNN, significantly reduced GPU memory, GPU process, CPU process, and virtual memory usage compared to traditional methods. For instance, UNet + FL + DNN required only 24.548% GPU memory (IID) and 27.975% GPU process (IID), a substantial reduction compared to MobileNetV2 (76.2% GPU memory) and UNet (69.5% GPU memory). The faster inference speeds (28 ms for UNet + FL + DNN on IID) further emphasize the practical deployability of these models in resource-constrained environments like small clinics or mobile diagnostic units. The adherence to regulatory frameworks like HIPAA and GDPR, combined with the slightly larger but manageable model sizes (27.1 MB for UNet + FL + DNN), makes these models highly suitable for decentralized clinical infrastructures, without compromising diagnostic accuracy or privacy.