Differential Privacy for Network Connectedness Indices
Novel two-layer privacy technique protects sensitive social network data while maintaining statistical accuracy.
A team of researchers including Tom A. Rutter, Yuxin Liu, and M. Amin Rahimian has developed a novel differential privacy method specifically designed for network connectedness indices. These indices measure assortative mixing—how people with similar attributes (like income, education, or political views) connect within social and economic networks. Traditional privacy techniques fail here because connectedness indices have high global sensitivity, and a single person's attribute can influence thousands of statistical cells, leading to poor privacy composition and excessive noise.
Their solution is a straightforward, two-step method: first adding calibrated noise to the sensitive node attributes themselves, then analytically debiasing the downstream network statistics, and finally applying a second layer of noise to protect the existence of individual connections (edges). The researchers prove their estimators are consistent and asymptotically normal for both discrete and continuous labels. Crucially, their method works effectively on real-world networks collected by social scientists with as few as 200 nodes, making it practical for actual research. All code to replicate their analyses is publicly available, facilitating adoption and verification.
- Method uses edge-adjacent differential privacy, a two-layer process protecting node attributes first, then individual edges.
- Proven effective on real social science networks with as few as 200 nodes, addressing a key practicality gap.
- Publicly released code allows researchers to safely analyze and publish sensitive network mixing patterns without privacy breaches.
Why It Matters
Enables safe, ethical analysis of sensitive social patterns—like economic segregation or political polarization—without exposing individual data.