Embedded Privacy-Aware Transformer With Sparse Attention for Consumer Health Monitoring

With the increasing integration of artificial intelligence into consumer healthcare devices, preserving the privacy and integrity of sensitive medical data has become paramount. Traditional deep learning models often overlook the privacy of the input level, leading to a possible exposure of sensitive attributes during training and inference. This paper presents an Embedded Privacy-Aware Transformer with Sparse Attention, specifically designed for real-time health monitoring on consumer healthcare platforms. The proposed model incorporates differential privacy through Laplace noise injection to anonymize sensitive input features, while efficiently processing nonsensitive features using an entropy-guided sparse attention mechanism. This design not only ensures robust privacy protection but also reduces computational overhead, making it suitable for deployment in resource-constrained consumer devices such as wearables and home health monitors. Experimental evaluation in a five-fold cross-validation lung cancer dataset demonstrates the effectiveness of the model, achieving an AUC of 0.982 and an F1 score of 0.95, while maintaining high utility and privacy guarantees in edge-enabled health monitoring scenarios.