Core Concepts
To achieve edge-differential privacy, graph coloring algorithms must have bounded defectiveness, where each vertex can share a color with at most a limited number of its neighbors.
Abstract
The paper studies the problem of vertex coloring in the differentially private setting. It shows that to be edge-differentially private, a coloring algorithm needs to be defective, where a vertex can share a color with at most a limited number of its neighbors. The authors prove a lower bound on the defectiveness of any differentially private c-edge coloring algorithm for graphs of maximum degree Δ, showing that the defectiveness must be at least Ω(log n / (log c + log Δ)).
The authors then present an ϵ-edge differentially private algorithm that produces an (O(Δ / (log n + 1/ϵ)), O(log n))-defective coloring, which is asymptotically tight for constant ϵ and Ω(log n) defectiveness. The algorithm first privately estimates the maximum degree of the graph, then augments the graph to ensure all vertices have approximately the same degree, and finally uses a random coloring approach to achieve the desired defectiveness.
Stats
The L1-sensitivity of the maximum degree is 1.
The additive error in estimating the maximum degree is bounded by α log n / ϵ with high probability.
Quotes
"To be edge-differentially private, a colouring algorithm needs to be defective: a colouring is d-defective if a vertex can share a colour with at most d of its neighbours."
"We show the following lower bound for the defectiveness: a differentially private c-edge colouring algorithm of a graph of maximum degree Δ> 0 has defectiveness at least d = Ω(log n / (log c + log Δ))."