Abstract: Privacy protection is important in visualization due to the risk of leaking personal sensitive information. In this paper, we study the problem of privacy-preserving visualizations using differential privacy, employing biomedical data from neuroimaging as a use case. We investigate several approaches based on perturbing correlation values and characterize their privacy cost and the impact of pre- and post-processing. To obtain a better privacy/visual utility tradeoff, we propose workflows for connectogram and seed-based connectivity visualizations, respectively. These workflows successfully generate visualizations similar to their non-private counterparts. Experiments show that qualitative assessments can be preserved while guaranteeing privacy. These results show that differential privacy is a promising method for protecting sensitive information in data visualization.
Introduction
Neuroimaging data is a double-edged sword: while it is incredibly rich for discovering biomarkers of diseases like Schizophrenia or Alzheimer’s, it is also highly sensitive. Sharing "raw" visualizations—such as heatmaps of brain activity or connectivity matrices—can inadvertently leak private information through "linkage attacks," where an attacker combines the visualization with external data to identify a specific individual.
Because of these risks, many researchers keep their data in "silos," leading to small sample sizes that lack statistical power. While Differential Privacy (DP) is the gold standard for protecting numerical data, applying it to visualizations is complex. Standard DP noise can turn a clear brain map into "salt-and-pepper" nonsense, destroying the visual utility that experts need to make sense of the data.
Methodology
The authors propose a multi-stage pipeline to balance the privacy-utility tradeoff, focusing on preserving the qualitative insights experts draw from neuroimaging data.
Sensitivity Reduction and Clipping
A critical step in the pipeline is reducing the sensitivity of the data. The authors use a data-driven clipping method to limit the range of correlation values. By narrowing the range (e.g., from [-1, 1] to [-0.75, 0.75]), they reduce the sensitivity, allowing for the addition of less noise to achieve the same level of privacy.
Differential Privacy Mechanisms
The study compared several noise-injection strategies, finding that the Analytic Gaussian mechanism generally performed better in high-dimensional brain data because it scales more favorably with the number of connections. They also explored matrix-level versus entry-level perturbation to optimize the preservation of the connectivity structure.
Specialized Workflows
The framework introduces specific workflows for common neuroimaging visualizations:
- Connectograms: For group comparisons (e.g., Healthy vs. Schizophrenia), a non-parametric Kruskal-Wallis test is integrated into the DP framework to identify significant brain regions.
- Seed-Based Connectivity (SBC): This workflow uses the Wilcoxon signed-rank test to identify significant voxels while protecting the underlying z-scores.
Post-processing and Denoising
Since DP is immune to post-processing, the authors used Singular Value Decomposition (SVD) and Haar Wavelet Transforms to "denoise" the visualizations. Because brain connectivity has a low-rank structure, SVD effectively filters out high-frequency noise without losing essential connectivity patterns.
Datasets
The study validated its methods across four major neuroimaging datasets:
- FBIRN (Function Biomedical Informatics Research Network): 1,056 subjects.
- BLSA (Baltimore Longitudinal Study of Aging): 891 subjects.
- ADNI (Alzheimer’s Disease Neuroimaging Initiative): 1,091 subjects.
- NITRC: Used specifically for evaluating Seed-Based Connectivity maps.
Key Results and Insights
- Qualitative Preservation: In connectogram experiments, the DP-protected visualizations successfully highlighted the same abnormalities in the sub-cortical and sensorimotor domains for Schizophrenia patients as the non-private versions.
- Quantitative Performance: The Approx-Clip-SVD method (Gaussian noise + clipping + SVD) consistently achieved the lowest Mean Squared Error (MSE) across all datasets and privacy budgets.
- Visual Utility vs. Accuracy: The authors argue that a private visualization is "useful" if an expert draws the same conclusion they would from the non-private version, even if the underlying numerical values have been perturbed.
Conclusion and Future Work
This research demonstrates that Differential Privacy is a promising method for protecting sensitive information in data visualization. While the current framework is robust, future work will involve human-in-the-loop studies to better quantify visual utility and the integration of these tools into federated learning systems like COINSTAC, allowing for global collaborative research without centralized data sharing.
Summary updated