To search, Click below search items. |
| All Published Papers Search Service |
|
|
|
Title |
Adaptive Anonymization for Privacy-Preserving Machine Learning: A Dynamic Approach to Secure Data Modeling |
|
Author |
Waad Saud AlAnazi and Jalal Suliman Alowibdi |
| Citation |
Vol. 26 No. 5 pp. 17-26 |
|
Abstract |
The fast proliferation of machine learning (ML) in data sensitive areas like healthcare, finance and cybersecurity has heightened the existence of privacy pre-serving strong privacy-mechanisms. The conventional anonymization methods, including k- anonymity, l-diversity, and t-closeness, fail to provide a balance be-tween privacy and data utility, resulting in poor model performance or the attack of inference. In response to these shortcomings, this paper will present an anonymization architecture that is more adaptive and changes the privacy levels dynamically based on the sensitivity of data and model needs. The proposed method is also learned with optimal anonymization parameters unlike the existing generalization-based methods, which rely on the concept of static generalization without data confidentiality. Experimental assessments with real-world samples and a wide variety of ML algorithms indicate that adaptive anonymization can be evaluated at 12% accuracy better than traditional solutions without decreasing the ability to resist linkage and attribute disclosure attacks. Comparative performances have shown that deep learning models are especially resistant to adaptive anonymization, maintaining more than 93 percent F1-score on anonymized data. These findings indicate that adaptive anonymization provides a privacy-compliant, scalable and task-aware method of data modeling to build a more trustful, transparent and resilient privacy preserving machine learning system. |
|
Keywords |
Adaptive anonymization; Privacy-preserving machine learning; Differential privacy; Cybersecurity; K-Anonymity |
|
URL |
