1 of 19

Data Sharing with Endogenous Choices over Differential Privacy Levels

Diptangshu Sen (Georgia Tech)

Joint work with Raef Bassily (OSU), Kate Donahue (MIT/UIUC),

Annuo Zhao, Juba Ziani (Georgia Tech).

16 June 2026

ESIF Economics and AI+ML Meeting || The Econometric Society

Cornell University, Ithaca NY

2 of 19

Motivation: data sharing

………

individually held data

largely insufficient!

share/pool data,

compute together!

Institutions need huge data to build powerful models and computations.

Example: hospitals trying to understand prevalence of new disease in population

Hospital A

Hospital B

Hospital X

sensitive patient data

solution?

3 of 19

Challenges

Privacy protections

many tools from differential privacy, but?
challenge – heterogenous privacy preferences

-> misaligned data sharing incentives among data owners!

Push to Decentralize Data Sharing

data owners want to retain control over data due to lack of trust
challenges – can lead to inefficient sharing!

4 of 19

This talk

Can large data sharing cooperatives be sustained under fully decentralized mechanisms with differential privacy?
Under what conditions?
How efficient are they compared to centralized mechanisms?

5 of 19

Differential Privacy 101

1-neighboring datasets x, x’

outcome distributions almost indistinguishable

Important property:

Post-processing immunity!

6 of 19

Setting

Target Population

No participation

= no benefits!

Privacy-accuracy tradeoff!

7 of 19

Types of Mechanisms

More Autonomy

More Efficiency

PoS

8 of 19

Accuracy (Variance) of Shared Estimator

Larger coalitions (higher |S|) better for accuracy!

Not clear if large coalitions can be sustained!

9 of 19

Social Cost

Each player incurs a burden or cost.
Player cost = Accuracy Cost + Privacy Cost

privacy preference for player ‘i’ (different for different players)

DP attack/observation model: how players perceive risk

10 of 19

What is f(S)? (1)

11 of 19

What is f(S)? (2)

Secure�Aggregation

“Get more privacy by hiding in the crowd”

12 of 19

What is f(S)? (3)

Leak (Economic Interpretation): “more people who know about me means more people with strategic power against me”

13 of 19

Full Decentralization: Stable Participation

14 of 19

Flavor of Results (Large ‘n’ regime)

15 of 19

Comparison

Centralized

Decentralized

PoS

Standard DP

Privacy leakage

Privacy amplfn

We provide a comparative analysis in terms of the privacy-cost regime (i.e., parameter \alpha) between fully centralized and fully decentralized settings. The green regions indicate regimes where we obtain non-trivial outcomes under either mechanisms. One immediate observation is that full decentralization offers non-trivial outcomes for strongly negative \alpha, i.e., when we have a high degree of privacy amplification due to data aggregation. Under the standard DP model (alpha = 0) or the leakage model (alpha = 0), the gains from decentralization are negligible. This is also reflected in the PoS, which becomes worse as we go from right to left. Another interesting observation is that for \alpha \in (-1/2, -1), while the grand coalition is optimal under both mechanisms and provide non-trivial outcomes, there is still a large efficiency gap which grows in the number of players to the 2/3. This shows that this efficiency gap is not due to stability or players being strategic about participation. But it is due to how the privacy levels are being chosen in either mechanism. Under full decentralization, selfish players consistently choose more stringent privacy levels (by a factor of k^1/3 where k is the coalition size) because they prioritize minimizing their own privacy cost as compared to maximizing the utility of the shared computation.

16 of 19

Key Takeaways

Data sharing incentives depend on the observation model!
Full decentralization is bad in general!
Useful only when players get strong privacy amplification due to aggregation.
Even when it is useful, it is very inefficient in practice!
Inefficiency not because of stability requirements, rather

because fully autonomous players choose privacy levels sub-optimally!

17 of 19

Thank You! Questions?

18 of 19

Data Sharing with Endogenous Choices over Differential Privacy Levels

Diptangshu Sen

Joint work with Raef Bassily, Kate Donahue, Annuo Zhao, Juba Ziani.

16 June 2026

ESIF Economics and AI+ML Meeting || The Econometric Society

Cornell University, Ithaca NY

19 of 19

Optimal Stable Coalitions at Large ‘n’

Interested in large coalitions which are:

stable;
optimal (social cost-minimizing out of all stable ones)

Properties:

variance of shared estimator
social cost of coalition

When non-trivial compared to no sharing?

No Data Sharing