Privacy-Preserving Active Learning for Medical Data Labeling Using Encrypted Uncertainty Measures

Abstract

Active learning (AL) significantly reduces annotation costs in medical applications by selecting the most informative unlabeled samples for expert labeling, potentially reducing required labels by 40–70%. However, centralizing sensitive healthcare data for uncertainty estimation and sample selection creates substantial privacy risks and regulatory compliance violations (HIPAA, GDPR). While Fully Homomorphic Encryption (FHE) enables machine learning on encrypted data without decryption, standard active learning query strategies cannot be directly applied due to computational overhead (100–1000× slower than plaintext) and the challenge of computing uncertainty measures directly on encrypted predictions. This paper introduces EALF, a novel framework combining active learning with fully homomorphic encryption for privacy-preserving medical data labeling in Internet of Medical Things (IoMT) systems. Our contributions include: (1) Encryption-aware uncertainty estimation techniques that compute Shannon entropy and margin-based measures directly over FHE-encrypted predictions, reducing computational overhead from ~100× to 4–9× compared to general FHE; (2) Genetic algorithm-based query strategy optimization that evolves domain-specific acquisition functions balancing uncertainty, representativeness, and misclassification costs in encrypted domains; (3) Federated encrypted active learning protocol enabling multi-institutional collaboration while maintaining patient privacy and institutional autonomy; and (4) Formal differential privacy analysis proving privacy preservation under realistic threat models. Comprehensive empirical evaluation on cardiac IoT datasets demonstrates that EALF achieves 70–76% reduction in required labeled samples with <500ms clinical latency while maintaining formal ε-differential privacy guarantees and robust resistance to membership inference attacks.