Anonymisation des données par apprentissage profonde(Data anonymization using deep learning)

Abstract

The increasing use of Electronic Health Records (EHRs) has created new opportunities for medical research and artificial intelligence, while raising significant concerns about patient privacy. Consequently, data anonymization has become essential for enabling secure data sharing without compromising the confidentiality of sensitive information. Traditional anonymization techniques, such as suppression, generalization, and differential privacy, often struggle to preserve the utility of complex longitudinal medical data. As an alternative, synthetic data generation has emerged as a promising anonymization approach, where artificial patient records are generated instead of releasing real ones, thereby preserving both privacy and data utility. Among deep learning methods, Generative Adversarial Networks (GANs) have shown great potential for synthetic medical data generation. However, existing GAN-based approaches suffer from two major limitations: mode collapse, which reduces the diversity of generated samples, and their black-box nature, which fails to explicitly model patients’ physiological states. To address these challenges, this thesis proposes a novel anonymization framework based on a Differentiable Hidden Markov Model (dHMM)- guided GAN. The proposed model combines explicit physiological state modeling with an LSTM-based patient history tracker to improve temporal consistency, while a MetropolisHastings-inspired latent space exploration strategy enhances sample diversity and mitigates mode collapse. Experimental results demonstrate that the proposed approach generates realistic, diverse, and clinically coherent synthetic medical records, providing an effective solution for privacy-preserving healthcare data sharing