ScholarWorks@숙명여자대학교

초록

Consumer healthcare devices generate continuous sensitive physiological data, enabling personalized medical insights, but expose users to sophisticated adversarial inference attacks. Traditional privacy mechanisms prove inadequate against intelligent adversaries adapting strategies based on observed data characteristics. We propose HealthGuard, a privacy-preserving framework protecting streaming healthcare data while preserving clinical utility through a dual-component architecture combining device-side intelligent preprocessing with server-side adversarial-resistant reconstruction. The framework introduces temporal-scale adaptive randomization, dynamically adjusting privacy budgets based on physiological significance, allocating stronger protection to rapid changes containing sensitive health information. Experimental validation on PAMAP2 and WESAD datasets comparing HealthGuard against eight baseline methods demonstrates membership inference attack reduction to an 8.7 percent success rate, a mean relative error of 0.065, 94.2 percent utility retention, and 62 percent lower computational overhead. Device-side measurements show 2 milliseconds of latency and 38 millijoules of energy per 1000 samples, enabling wearable deployment. Scalability analysis demonstrates sublinear growth supporting 5000 devices with 42 percent overhead reduction. Cross-domain evaluation yields a 1.5 percentage point degradation, validating transferability across heterogeneous consumer healthcare scenarios.

키워드