A Survey on Differentially Private Machine Learning [Review Article]

Abstract

R ecent years have witnessed re markable successes of machine learning in various applications. However, machine learning models suffer from a potential risk of leaking private information contained in training data, which have attracted increasing research attention. As one of the mainstream priva cypreserving techniques, differential pri vacy provides a promising way to prevent the leaking of individuallevel privacy in training data while preserving the quality of training data for model building. This work provides a comprehensive survey on the existing works that incorporate differ ential privacy with machine learning, socalled differentially private machine learning and categorizes them into two broad categories as per different differen tial privacy mechanisms: the Laplace/ Gaussian/exponential mechanism and the output/objective perturbation mecha nism. In the former, a calibrated amount of noise is added to the nonprivate model and in the latter, the output or the objective function is perturbed by ran dom noise. Particularly, the survey covers the techniques of differentially private deep learning to alleviate the recent con cerns about the privacy of big data con tributors. In addition, the research challenges in terms of model utility, priva cy level and applications are discussed. To tackle these challenges, several potential future research directions for differentially private machine learning are pointed out.

Multiple Strategies Differential Privacy on Sparse Tensor Factorization for Network Traffic Analysis in 5G

“…, when λ is the noise parameter [21]. This mechanism mainly consists of two steps: perturbation statistics and correction [7,14].…”

“…More recently, due to the appetency to provide a privacy guarantee [7], DP naturally becomes a preferred tool for data privacy protection in the training process of Machine Learning (ML) models [8][9][10]. DP can provide privacy protection for the classic dimension reduction model, i.e., Matrix Factorization (MF) [11,12] and the application fields have been extended to recommender systems and social networks [13,14].…”

Adversarial Attacks Against Deep Generative Models on Data: A Survey

“…Serban et al elaborated on adversarial examples, including their construction, defense strategies, and transfer capabilities [26]. As differential privacy is one of the most effective measure for mitigating privacy breaches, Gong et al published a comprehensive review on differentially-private machine learning [27]. A number of surveys have also been conducted on DDMs, particularly CNNs and RNNs, see e.g., [28], [29], [30], [31], [32].…”

FL-MAC-RDP: Federated Learning over Multiple Access Channels with Rényi Differential Privacy

“…In order to improve the communication efficiency of FL-DP, Sonne and Rin [10] study FL over MAC with DP (FL-MAC-DP). However, the traditional DP have some limitations [29] in calculating the cumulative privacy loss. Then Rényi differential privacy (RDP) is proposed in [30], as a natural relaxing of traditional DP, it is suitable to express the privacy protection algorithm and the composition of heterogeneous mechanism.…”