{"paper":{"title":"Differentially Private Model Merging","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Post-processing existing differentially private models with random selection or linear combination produces a model meeting any target privacy level without additional training.","cross_cats":["cs.AI","cs.CR","stat.ML"],"primary_cat":"cs.LG","authors_text":"Manzil Zaheer, Qichuan Yin, Tian Li","submitted_at":"2026-04-22T18:13:37Z","abstract_excerpt":"In machine learning, privacy requirements at inference or deployment time often evolve due to changing policies, regulations, or user preferences. In this work, we aim to construct a magnitude of models to satisfy any target differential privacy (DP) requirement without additional training, given a set of existing models trained on the same dataset with different privacy/utility tradeoffs. We propose two post-processing techniques, namely random selection and linear combination, to generate final private models satisfying any target privacy parameter. We provide privacy accounting of these app"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We propose two post processing techniques, namely random selection and linear combination, to output a final private model for any target privacy parameter. We provide privacy accounting of these approaches from the lens of Rényi DP and privacy loss distributions for general problems. In a case study on private mean estimation, we fully characterize the privacy/utility results and theoretically establish the superiority of linear combination over random selection.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The existing models are trained on the same dataset with different privacy/utility tradeoffs, which is required for the post-processing techniques to correctly compose privacy guarantees without additional training.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Post-processing via random selection or linear combination generates differentially private models for arbitrary privacy parameters from pre-trained models on the same dataset.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Post-processing existing differentially private models with random selection or linear combination produces a model meeting any target privacy level without additional training.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"f4dd4840f1b51e0df8bb48daed3b13592d99603956ffab65f9375a43a5438bac"},"source":{"id":"2604.20985","kind":"arxiv","version":2},"verdict":{"id":"4ed4ce60-10a2-4ff3-a88e-1782a20a0e27","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-10T01:10:56.130853Z","strongest_claim":"We propose two post processing techniques, namely random selection and linear combination, to output a final private model for any target privacy parameter. We provide privacy accounting of these approaches from the lens of Rényi DP and privacy loss distributions for general problems. In a case study on private mean estimation, we fully characterize the privacy/utility results and theoretically establish the superiority of linear combination over random selection.","one_line_summary":"Post-processing via random selection or linear combination generates differentially private models for arbitrary privacy parameters from pre-trained models on the same dataset.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The existing models are trained on the same dataset with different privacy/utility tradeoffs, which is required for the post-processing techniques to correctly compose privacy guarantees without additional training.","pith_extraction_headline":"Post-processing existing differentially private models with random selection or linear combination produces a model meeting any target privacy level without additional training."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.20985/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_compliance","ran_at":"2026-05-20T01:29:53.754074Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"482d38c29369652c0ab2d1fc2260afb03061e5dbcdd5312859ad9fa204dff060"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}