An Efficient Framework for Detecting Model Inversion Attacks in Connected Healthcare Systems

초록

The rapid adoption of artificial intelligence (AI) in healthcare systems and consumer electronics is enhancing clinical decision-making and reshaping patient care. However, these advancements also raise serious security and privacy concerns, particularly model inversion attacks (MIAs), which reconstruct sensitive patient data from trained AI models. Addressing MIAs is essential to ensure trustworthiness and effectiveness of AI-driven healthcare solutions. To address these challenges, a novel four-stage detection framework is proposed to safeguard healthcare AI models against MIAs. Firstly, the data preprocessing and normalization layer eliminates the input noise and stabilizes the incoming query patterns. Onward, the behavioral feature extraction is carried out. It seeks to detect adversarial intent by profiling the temporal patterns as well as the irregularities found in the query behavior. Afterward, statistical anomaly scoring is applied using Mahalanobis distance learning to quantify deviations in query behavior. Lastly, an ensemble-based decision module, built on support vector machine and random forest classifiers, is used for the identification of MIAs. It combines anomaly scores and utilizes a threshold-based decision to accurately separate benign and possible MIA sessions while enhancing robustness against sophisticated attacks. Extensive experimental results on the NIH ChestXray14 and MIT-BIH datasets show that the proposed framework improves best-performing benchmarks in terms of precision by 1.5%, recall by 2.8%, F1-score by 1.18%, and AUROC by 2.23%, while reducing overhead by 9.09%.

키워드