Enabling Humanitarian Applications with �Targeted Differential Privacy*

Nitin Kohli

MeasureDev 2026 – April 29, 2026

Staff Scientist, UC Berkeley Center for Effective Global Action

*Joint work with Professor Joshua Blumenstock

2

Novissi in Togo

TOGO

• Digital Emergency Cash Transfer Program
• “Poorest individuals in the poorest regions”
• $13-15 per month for 5 months
• Launched: April 2020
• Reach: 800K individuals

​

• Challenge: Targeting transfers without administrative data

Two Stage Approach

3

Extended Data Figure 1A: Aiken et al. (2022) Nature.

Stage 1: Identify Poorest Regions

Stage 2: Identify Poorest Individuals

Source: Aiken et al. (2022) Journal of Development Economics; Figure 4 Supplemental Information

Poor

Non-poor

Validation of Step 2 of Original Novissi on Older Data (2018)

4

Figure 1B: Aiken et al. (2022)

5

Issue: Mobile phone metadata can reveal sensitive information

Fig. 2 in de Montjoye et al. (2013); Scientific Reports

Provable privacy meets humanitarian targeting

• Our technical work is enabling new use cases
• Develop new privacy definitions and algorithms that provide strict privacy guarantees, while also facilitating downstream reuse, and applications of machine learning

​

​

​

​

2. Characterizing the real-world tradeoffs induced by increasing privacy

A: Non-Private Data

(Owned by Telco)

C: Privatized Data

(Provided to gov’t, bank, researchers, etc.)

D: Labels

Details: Kohli and Blumenstock (2024), “Enabling Humanitarian Applications with Targeted Differential Privacy”

Candidate Approach: Differential Privacy (DP)

• Well-suited for “classic statistics”
• Goal:
• Learn about a population,
• Without learning “too much” about any individual
• Approach:
• Add carefully-calibrated noise to mask the contribution of any individual datapoint

DB with your data as x

Randomized Algorithm

Synthetic Dataset

DB with your data as x’

With

“essentially the same” probability

Intuition

Adversary cannot discern

individual’s true data

But “classic stats” approaches are ill-suited for targeting

Poor

Non-Poor

Randomized Computation

With

“essentially the same” probability

Difficult for

ML to

learn

target

• Well-suited for targeting where individual learning is requiring
• Goal:
• Enable population level learning,
• And “some” individual learning,
• But “not too much”
• Approach:
• Add carefully-calibrated noise to mask the value of “sufficiently similar” individuals

DB with your data as x

Private Algorithm

Output

DB with your data as x’

With

“essentially the same” probability

�Targeted

Differential Privacy

Intuition

Adversary can discern between “dissimilar” inputs but not confidently for “similar” inputs

Contribution 1a: Targeted Differential Privacy

Necessary condition for accurate targeting

10

Contribution 1b: Projection Algorithm

D: Targets

B: Private Projection Algorithm

E: Machine Learning

Private Projection Algorithm (details)

The Targeting Setting

…

R₍₁₎

R₍₂₎

R_(k)

F: Applications

G: Individual i’s personal data

Contribution 1: TDP + Projection Algorithm

Our algorithm for targeted differential privacy substantially outperforms differential privacy, when evaluated using Togo data

​

Accuracy of humanitarian aid allocation using different PETs

115K and 240K additional

exclusion errors

2K additional exclusion errors

No Privacy

Differential

Privacy

k-Anonymity

Targeted Diff.

Privacy

Contribution 2: Privacy-Program Effectiveness Tradeoff

• Singling-out attacks

• Concern: Can an adversary isolate at least one individual in the real data (based on unique combinations of features)?

​

Status quo:

No provable

privacy protections

Strong protections,

minor exclusion error

Strongest protections,

untenable exclusion error

No Privacy

Targeted Differential Privacy

Differential Privacy

Targeting of Anti-Poverty Assistance in Togo

Takeaway

• Humanitarian applications are increasingly relying on sensitive data for interventions
• For targeting applications, existing PETs are misaligned with the goals of targeting (i.e., individual-level learning)
• Targeted differential privacy provides a contextual adaptation of differential privacy to facilitate humanitarian applications with strong privacy guarantees

Appendix Slides

17