A unified framework for secure and intelligent ECG signal processing via chaotic encryption, fully homomorphic computation and federated learning

Abstract

The digital transformation of healthcare is changing how cardiovascular diseases are detected, monitored, and treated. Continuous electrocardiogram (ECG) monitoring plays an important role in early diagnosis and preventive care. However, as patient data is increasingly collected through connected devices and cloud-based systems, data security and privacy have become major concerns. There is a clear need for systems that can securely collect, transmit, store, and analyze physiological signals without reducing diagnostic accuracy or clinical usability. This doctoral research addresses this need by proposing an integrated privacy-preserving ECG monitoring and analysis system that combines real-time signal acquisition, encryption, and distributed artificial intelligence. The thesis consists of three related studies, each focusing on a different aspect of secure biomedical data processing. The first study investigates real-time privacy-preserving ECG monitoring, the second study introduces an adaptive chaotic encryption method, and the third study examines collaborative model training using federated learning under encryption. Together, these studies form a unified system that protects ECG data at every stage, from acquisition to computation, while maintaining real-time performance required in clinical environments. The first study focuses on the development of a real-time ECG monitoring system with built-in privacy protection. The hardware is based on a three-lead ECG preamplifier connected through a serial interface, allowing continuous and reliable signal acquisition. The software platform is implemented in Python and performs three main tasks: real-time ECG monitoring and disease detection, secure cloud storage using Advanced Encryption Standard (AES), and statistical analysis on encrypted data using Fully Homomorphic Encryption (FHE). This approach allows ECG data to be processed in the cloud without being decrypted. Authorized medical staff can remotely access, decrypt, and visualize the data while preserving synchronization between local and cloud systems. Experimental evaluations showed that encryption and transmission delays remained below 50 milliseconds per segment, confirming that real-time clinical operation is possible without loss of diagnostic accuracy. The second study focuses on signal-level data security through a Learnable Chaotic Encryption (LCE) method. Unlike traditional chaotic encryption schemes that use fixed parameters, the proposed method generates encryption keys dynamically from the ECG signal itself. A Learnable Key Generator estimates the logistic map parameters r and x₀ using the statistical properties of each ECG segment, specifically the mean and standard deviation. These parameters are used to create a unique chaotic sequence for each segment. The sequence controls permutation and XOR-based diffusion steps, making each encrypted segment unique and closely linked to the patient's physiological signal. This design increases resistance to prediction, brute-force attacks, and unauthorized reconstruction. Extensive security evaluations confirmed the effectiveness of the proposed encryption method. Shannon entropy values ranged between 7.6 and 7.8 bits, and correlation values were close to zero, indicating strong randomness and low information leakage. Avalanche effect analysis showed more than 55% bit change when a single key bit was modified, demonstrating high sensitivity to key variation. Randomness was further validated using the NIST SP800-22 test suite. Decryption results showed near-lossless signal recovery with peak signal-to-noise ratio values above 52 dB and very low mean squared error, confirming that diagnostically important signal features were preserved. The encryption process operated within real-time limits, making it suitable for wearable devices and Internet of Medical Things applications. The third study extends the secure ECG processing system to a collaborative learning setting using Federated Learning combined with Homomorphic Encryption. The goal is to allow multiple hospitals or research institutions to train a shared diagnostic model without exchanging raw patient data. The PTB-XL ECG dataset is used and divided into horizontal, vertical, and combined scenarios to represent different clinical data distributions. Each institution trains a local deep learning model on its own data and sends only encrypted gradient updates to a central aggregation server. Homomorphic encryption ensures that the transmitted gradients cannot be analyzed to recover sensitive data, even if the server is compromised or operates under an honest-but-curious assumption. In our experiments, the aggregated global model achieved diagnostic accuracies of up to 99%, which were very close to those obtained using centralized learning. Communication cost varied across the different federated learning scenarios, with horizontal data partitioning providing the best balance between training speed and security. These results showed that large-scale collaborative medical model training is possible under strict privacy requirements and in compliance with data protection regulations such as GDPR and KVKK. Overall, the three studies together form a complete privacy-preserving ECG analysis system. The first study ensures secure real-time data acquisition and transmission, the second study provides adaptive and patient-specific signal encryption, and the third study enables secure collaborative learning across institutions. By combining these components, the system ensures that ECG data remains protected during storage, transmission, and computation. Beyond technical contributions, this thesis demonstrates that strong data protection can be achieved without reducing clinical performance. The proposed system supports real-time operation, meets regulatory requirements, and maintains high diagnostic accuracy. All software components and experimental tools developed in this research are released as open-source resources to support transparency and reproducibility. In conclusion, this dissertation shows that future healthcare systems must be designed to be both intelligent and privacy-aware. By integrating privacy directly into data acquisition, processing, and learning stages, this work provides a practical and reliable approach for secure medical artificial intelligence applications.