Differentially Private Model Merging via Post-Processing

A new arXiv preprint (2604.20985) proposes two post-processing techniques—random selection and linear combination—to generate models satisfying any target differential privacy (DP) requirement without additional training. The methods leverage existing models trained on the same dataset with varying privacy/utility tradeoffs. Privacy accounting is provided using Rényi DP and privacy loss distributions. A case study on private mean estimation shows linear combination outperforms random selection both theoretically and empirically.