Abstract
This paper analyzes the vulnerabilities of privacy-preserving algorithms applied to multidimensional medical and nutritional databases. Syntactic anonymization models (k-anonymity, l-diversity, t-closeness), differential privacy and federated learning are compared. It is shown that each of them is exposed to specific attacks: homogeneity and semantic similarity attacks, privacy budget exhaustion under composition, and gradient inversion. The need for adaptive, dynamic privacy models is substantiated.
References
[1] Sweeney L. k-anonymity: a model for protecting privacy // International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems. – 2002. – Vol. 10, No. 5. – P. 557–570.
[2] Machanavajjhala A., Kifer D., Gehrke J., Venkitasubramaniam M. l-diversity: privacy beyond k-anonymity // ACM Transactions on Knowledge Discovery from Data. – 2007. – Vol. 1, No. 1. – Article 3.
[3] Li N., Li T., Venkatasubramanian S. t-closeness: privacy beyond k-anonymity and l-diversity // Proc. IEEE 23rd International Conference on Data Engineering (ICDE). – 2007. – P. 106–115.
[4] Dwork C., Roth A. The algorithmic foundations of differential privacy // Foundations and Trends in Theoretical Computer Science. – 2014. – Vol. 9, No. 3–4. – P. 211–407.
[5] McMahan H.B., Moore E., Ramage D., Hampson S., Agüera y Arcas B. Communication-efficient learning of deep networks from decentralized data // Proc. AISTATS. – 2017. – P. 1273–1282.
[6] Zhu L., Liu Z., Han S. Deep leakage from gradients // Advances in Neural Information Processing Systems 32 (NeurIPS). – 2019. – P. 14774–14784.
[7] Abadi M. et al. Deep learning with differential privacy // Proc. ACM CCS. – 2016. – P. 308–318.