Go to DBLP homepageSecure Sleep Apnea Detection with FHE and Deep Learning on ECG Signals
Abstract: Sleep apnea, a prevalent sleep disorder affecting individuals of all demographics, poses a threat of significant disruption to daily life. The analysis of Electrocardiogram (ECG) data facilitates the accurate diagnosis of sleep apnea. With the advent of machine learning and its accessibility through cloud services, doctors have been compelled to enhance their diagnostic capabilities by integrating deep learning into their analytical tools. However, challenges such as data privacy, security, and confidentiality regulations are hindering the adoption of deep learning in the healthcare domain. In this research, we address these challenges by proposing an end-to-end encrypted framework to analyze encrypted ECG signals and diagnose sleep apnea. Leveraging Fully Homomorphic Encryption (FHE) on deep learning models ensures privacy and security by design while enabling computations on encrypted data. To overcome the unique challenges posed by handling encrypted data in deep learning models, we introduce novel and efficient techniques for adapting several key components such as the convolutional layer, max pooling, ReLU activation, and fully connected layer to the FHE domain. Our approach includes adapting the convolutional layer in the spectral domain, implementing fully connected layers as generalized matrix multiplication, and employing approximation methods for ReLU activation and max pooling. The experimental results on real encrypted ECG data demonstrate the feasibility and efficacy of our proposed framework, achieving a remarkable accuracy of 99.56% in detecting sleep apnea. Our proposed encrypted network does not lose any predictive performance compared to its plaintext counterpart. This research underscores the potential of encrypted data processing in significantly enhancing the security and privacy of healthcare services, particularly in the domain of sleep apnea diagnosis.
Loading