Federated learning boosts skin disease detection without compromising privacy

The proposed federated learning system enables model training directly on edge devices, sending only encrypted model updates to a central server for aggregation. This decentralization eliminates the need for raw data transfer, significantly reducing vulnerability.

Traditional skin disease diagnostic systems rely heavily on centralized machine learning architectures that require sensitive patient images to be transmitted to central servers for training. This exposes data to privacy risks, including breaches and non-compliance with GDPR and HIPAA regulations. To address this issue, researchers have developed a new federated learning framework that shows exceptional accuracy in classifying skin diseases without compromising sensitive patient data - a breakthrough for privacy-conscious healthcare.

Published in Frontiers in Computer Science, the study titled “Privacy-enhanced skin disease classification: integrating federated learning in an IoT-enabled edge computing” introduces a decentralized AI model that achieves a remarkable 98.6% classification accuracy across nine dermatological conditions. The system leverages edge devices and Internet of Things (IoT) sensors to analyze images locally, ensuring that raw medical data never leaves the device.

How does the model ensure data privacy without sacrificing accuracy in skin disease diagnosis?

The proposed federated learning system enables model training directly on edge devices, sending only encrypted model updates to a central server for aggregation. This decentralization eliminates the need for raw data transfer, significantly reducing vulnerability.

The model uses InceptionV3 as its base classifier, capable of processing RGB images at a resolution of 299×299 pixels. Local model updates are aggregated on the central server using a secure federated averaging process enhanced with cryptographic protocols such as multiparty computation (MPC) and secure aggregation. These methods ensure that individual device updates remain private, preventing model inversion attacks and unauthorized data reconstruction. The system also incorporates adaptive model partitioning to distribute tasks based on each device’s computational capacity, ensuring minimal latency and efficient network load balancing.

To increase training diversity, the researchers used the ISIC dataset and supplemented it with samples from HAM10000 and DermNet, applying synthetic data generation to represent different ethnicities and age groups. Images were uniformly scaled using bilinear interpolation and subjected to augmentation techniques such as pixel translation and rotation to increase variability and improve generalization.

How are edge devices managed and evaluated within the federated architecture?

The model was deployed across four edge nodes, each configured with varying computational capabilities and data types. The data was probabilistically distributed using stochastic data distribution to ensure fairness and heterogeneity, a key factor in training robust machine learning models. To handle imbalances and ensure computational efficiency, edge devices with lower capacity were assigned lighter workloads, while stronger devices processed more complex data.

The network achieved high stability by shuffling datasets before training and using batch sizes of 32 images per iteration. The training spanned 14 epochs, and a central validation set was maintained for global evaluation. Classification accuracy, precision, recall, and F1-score were calculated per skin disease class. Performance scores were near perfect for most classes: Actinic keratosis, vascular lesions, and basal cell carcinoma scored F1-scores of 1.0, 1.0, and 0.99 respectively. Even the more complex classes such as dermatofibroma and squamous cell carcinoma achieved F1-scores of 0.96 and 0.95.

The framework also monitored training and validation loss curves to ensure the model avoided overfitting and underfitting. Convergence patterns across multiple edge nodes revealed consistent performance gains, confirming the system’s adaptability and reliability in distributed environments. Confusion matrices further validated class-specific accuracy and showed minimal cross-category misclassification, such as occasional overlap between melanoma and nevus.

What does this research mean for the future of privacy-first AI in dermatology and beyond?

The implications of this research extend far beyond skin disease classification. The proposed framework demonstrates that federated learning, when integrated with IoT and edge computing, can achieve hospital-grade diagnostic accuracy while fully complying with privacy regulations. This makes it highly scalable and adaptable for use in underserved regions, teledermatology services, and real-time diagnostic platforms embedded in wearable or mobile devices.

Energy efficiency and resource utilization were also assessed, showing the model's viability in real-world IoT environments. Using a Raspberry Pi 4 as a reference edge device, the system consumed only 5.6 watts during training, 3.2 watts during inference, and required just 2.2 MB per communication batch. These metrics confirm that the model can be deployed in low-resource settings without incurring significant energy or bandwidth costs.

Compared to benchmark models like SENet154, Ensemble CNN, and MobileNet, the proposed system delivered the highest overall performance across multiple metrics, including a precision of 0.99 and an F1-score of 0.98. By integrating secure aggregation and a collaborative training structure, it also addressed one of the most pressing barriers to AI deployment in healthcare: patient data protection.