When Privacy Moves ML-Mediated Decisions On Device: Information and Incentive Misalignment in Auctions
On-device auction simulation shows stale budgets cause large overspend and incentive-breaking payment deviations.
The paper studies on-device ML auctions where privacy keeps shared budgets from staying globally current. Across 36 campaigns and 50 devices, proportional Even pacing overspends 17.77% after one tick of staleness and 1,669.31% after 50 ticks under 20× budget pressure; at 2× pressure the 50-tick figure is still 106.95%. A visible-budget no-sale guard leaves 11.88% one-tick overspend, and 98.23% of rival auctions at one tick admit a profitable deviation when score transforms change payment units.
- Simulation uses 36 campaigns, 50 devices, and dimensionless score units.
- One-tick staleness causes 17.77% overspend; 50 ticks reach 1,669%.
- Two-times budget pressure still yields 106.95% overspend at 50 ticks.
- Payment changes make 98.23% of one-tick rival auctions manipulable.
From the original article
1 min read at huggingface.co
Moving ML-mediated decision making onto privacy-preserving clients decentralises the economic decision along with the inference. Shared budget constraints then depend on information that cannot be globally current, creating an information-structure failure that conventional pacing is not designed to solve.
By Dipankar Sarkar · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.