Multi-Agent Machine Unlearning in Healthcare AI: Neural, Symbolic, and Coordination

Abstract

The growth of AI systems in healthcare has led to complex ecosystems where the same patient information affects several interlinked models and knowledge bases. Current system unlearning techniques tackle some of those system components but do not guarantee complete data removal across entire healthcare structures. This paper presents a comprehensive integrated forgetting framework featuring neural unlearning, symbolic knowledge transferring and multi-agent consensus protocols. Our approach enables formal guarantees of privacy using differential privacy mechanisms and cryptographic verification. The framework is validated by two extensive simulation studies: synthetic datasets (over 100, 000 patients and 50 AI systems) and real dataset healthcare arranged scenarios (involving over 12,000 patient records across 8 clinical AI systems). The simulation experiment results imply that 84.7 percent of forgetting accuracy with <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\varepsilon$</tex>-differential privacy (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\varepsilon= 0.01$</tex>), 82.3 percent of utility retention and 3.2 seconds of average time to complete the deletion has been obtained. The simulated model is reachable to 35 % in the membership inference attack and has a reasonable robustness against reconstruction attack (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$M S E=0.73$</tex>) while maintaining HIPAA compliance. Clinical validation by simulation-based testing using 45 health professionals has demonstrated an acceptance rate of 94% signifying the deployment feasibility.

Ähnliche Arbeiten