Secure and Differentially Private Logistic Regression for Horizontally Distributed Data

Abstract

Scientific collaborations benefit from sharing information and data from distributed sources, but protecting privacy is a major concern. Researchers, funders, and the public in general are getting increasingly worried about the potential leakage of private data. Advanced security methods have been developed to protect the storage and computation of sensitive data in a distributed setting. However, they do not protect against information leakage from the outcomes of data analyses. To address this aspect, studies on differential privacy (a state-ofthe-art privacy protection framework) demonstrated encouraging results, but most of them do not apply to distributed scenarios. Combining security and privacy methodologies is a natural way to tackle the problem, but naive solutions may lead to poor analytical performance. In this paper, we introduce a novel strategy that combines differential privacy methods and homomorphic encryption techniques to achieve the best of both worlds. Using logistic regression (a popular model in biomedicine), we demonstrated the practicability of building secure and privacy-preserving models with high efficiency (less than 3 min) and good accuracy [<1% of difference in the area under the receiver operating characteristic curve (AUC) against the global model] using a few real-world datasets.

Truly Privacy-Preserving Federated Analytics for Precision Medicine with Multiparty Homomorphic Encryption

“…In contrast to the conventional federated learning approach, which requires each data provider to perturb its local results before aggregating them with other parties, FAMHE enables the data providers to keep the local results encrypted and reveals only the final aggregated results. Therefore, FAMHE can use a smaller amount of added noise and achieve the same level of privacy 38 . Notably, the choices of differential-privacy parameters suitable for analyses with a high-dimensional output, such as GWAS, can be challenging and needs to be further explored.…”

“…By keeping the intermediate results encrypted (and never decrypting them), FAMHE enables the perturbation to be applied on only the final result, i.e., before the Collective Key Switching. In this case, the DPs collectively generate the perturbating noise by generating it from composable distributions, e.g., Laplace, as explained by Kim et al 39 . As the function's output sensitivity is the same for the DPs' local results and for the combined final result, FAMHE enables a smaller (proportionally to the number of DPs) perturbation of the final result for the same level of privacy.…”

“…This is a notable improvement with respect to existing federated-learning solutions based on differential privacy 39 in which the noise has to be added by each data provider at each iteration of the training. In the solution by Kim et al 39 , each data provider perturbs its locally computed gradient such that the aggregated perturbation, obtained when the data providers aggregate (combine) their locally updated model, is e-differentially private. This is achieved by having each data provider generating and adding a partial noise such that, when aggregated, the total noise follows the Laplace distribution.…”

Forests Growth Monitoring Based on Tree Canopy 3D Reconstruction Using UAV Aerial Photogrammetry

“…The logistic S-curve application model is based on firmly proven laws of nature [24][25][26][27]. The S-curve model represents the growth or decline of every system in interaction with its environment (its limited resources).…”

Multicenter Privacy-Preserving Cox Analysis Based on Homomorphic Encryption

“…Many solutions have been developed to address privacy-preserving data analysis, including logistic regression [8,9], support vector machines [10,11], and linear regression [12,13]; however, only a few of these have focused on ensuring the security of Cox regression analysis. Yu et al (2008) first proposed a privacy-preserving Cox model training process in which they projected patient data to a lower-dimensional space and used this space to train the model [14].…”