Differential privateness (DP) is a mathematically rigorous and broadly studied privateness framework that ensures the output of a randomized algorithm stays statistically indistinguishable even when the information of a single person adjustments. This framework has been extensively studied in each principle and observe, with many functions in analytics and machine studying (e.g., 1, 2, 3, 4, 5, 6, 7).
The 2 important fashions of DP are the central mannequin and the native mannequin. Within the central mannequin, a trusted curator has entry to uncooked knowledge and is chargeable for producing an output that’s differentially non-public. The native mannequin requires that every one messages despatched from a person’s machine are themselves differentially non-public, eradicating the necessity for a trusted curator. Whereas the native mannequin is interesting because of its minimal belief necessities, it usually comes with considerably increased utility degradation in comparison with the central mannequin.
In real-world data-sharing situations, customers usually place various ranges of belief in others, relying on their relationships. As an illustration, somebody may really feel snug sharing their location knowledge with household or shut mates however would hesitate to permit strangers to entry the identical info. This asymmetry aligns with philosophical views of privateness as management over private info, the place people specify with whom they’re prepared to share their knowledge. Such nuanced privateness preferences spotlight the necessity for frameworks that transcend the binary belief assumptions of current differentially non-public fashions, accommodating extra sensible belief dynamics in privacy-preserving techniques.
In “Differential Privateness on Belief Graphs”, revealed on the Improvements in Theoretical Pc Science Convention (ITCS 2025), we use a belief graph to mannequin relationships, the place the vertices symbolize customers, and related vertices belief one another (see under). We discover the best way to apply DP to those belief graphs, guaranteeing that the privateness assure applies to messages shared between a person (or their trusted neighbors) and everybody else they don’t belief. Particularly, the distribution of messages exchanged by every person u or certainly one of their neighbors with another person not trusted by u ought to be statistically indistinguishable if the enter held by u adjustments, which we name belief graph DP (TGDP).