{"paper":{"title":"Jacobian-Guided Anisotropic Noise Reshaping for Enhancing Representation Utility under Local Differential Privacy","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Jacobian-guided reshaping turns isotropic LDP noise anisotropic to raise representation utility while holding the per-dimension privacy budget fixed.","cross_cats":["cs.CR"],"primary_cat":"cs.LG","authors_text":"Anil Anthony Bharath, Viktor Schlegel, Yidan Sun, Youngmok Ha","submitted_at":"2026-05-16T05:01:41Z","abstract_excerpt":"While Local Differential Privacy (LDP) serves as a foundational primitive for distributed data collection, its stringent noise injection requirement often leads to severe degradation in data utility. This degradation stems from the task-agnostic nature of conventional LDP mechanisms, which inject noise uniformly across all dimensions regardless of their relative importance to the downstream objective. To address this issue, we propose a novel approach that mitigates noise in task-relevant subspaces of the data representation. Our method identifies task-critical subspaces via the Jacobian matri"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Integrating our approach improves the utility of PrivUnit2 and PrivUnitG by approximately 20% at ε=7.5 on CIFAR-10-C (Brightness corruption at severity level 5) while preserving the uniform per-dimension privacy budget.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The Jacobian matrix of the public downstream model accurately identifies task-critical subspaces without introducing additional privacy leakage or requiring knowledge unavailable under the LDP threat model (abstract, section on method).","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Jacobian-guided anisotropic noise reshaping improves LDP utility by ~20% on CIFAR-10-C at ε=7.5 by attenuating noise in task-critical subspaces identified via the downstream model's Jacobian.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Jacobian-guided reshaping turns isotropic LDP noise anisotropic to raise representation utility while holding the per-dimension privacy budget fixed.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"8dfe7e9b8d9a7d5f52b34f86863a152434ef18c999adf06b67a2a9d2916b80dc"},"source":{"id":"2605.16812","kind":"arxiv","version":1},"verdict":{"id":"d53d68dd-0e44-447f-9a46-bd4a2fe5a4da","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T20:59:09.754176Z","strongest_claim":"Integrating our approach improves the utility of PrivUnit2 and PrivUnitG by approximately 20% at ε=7.5 on CIFAR-10-C (Brightness corruption at severity level 5) while preserving the uniform per-dimension privacy budget.","one_line_summary":"Jacobian-guided anisotropic noise reshaping improves LDP utility by ~20% on CIFAR-10-C at ε=7.5 by attenuating noise in task-critical subspaces identified via the downstream model's Jacobian.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The Jacobian matrix of the public downstream model accurately identifies task-critical subspaces without introducing additional privacy leakage or requiring knowledge unavailable under the LDP threat model (abstract, section on method).","pith_extraction_headline":"Jacobian-guided reshaping turns isotropic LDP noise anisotropic to raise representation utility while holding the per-dimension privacy budget fixed."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.16812/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_title_agreement","ran_at":"2026-05-19T21:31:19.259775Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T21:11:43.382026Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"claim_evidence","ran_at":"2026-05-19T19:01:56.276980Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T18:33:26.415956Z","status":"skipped","version":"1.0.0","findings_count":0}],"snapshot_sha256":"5c468dc09ff4f36fb561d1f112ddb8269fab265442480b40d4c9b19f34b7da55"},"references":{"count":78,"sample":[{"doi":"","year":2006,"title":"Theory of Cryptography Conference (TCC) , pages =","work_id":"d20b7570-d7a2-4e2f-807e-ce3d686924e5","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"SIAM Journal on Computing , volume =","work_id":"473c5452-1d91-400e-b6e2-619f0ff2a54b","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Proceedings of the ACM SIGSAC Conference on Computer and Communications Security (CCS) , pages =","work_id":"ff49c343-9167-4146-8231-661788f8277a","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Advances in Neural Information Processing Systems (NIPS) , pages =","work_id":"ede9fb2e-e4b5-48dd-9b3f-61ef80f08ddd","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"ICML Workshop on Federated Learning and Analytics in Practice , year =","work_id":"72b5b59d-4913-4a47-87c0-23d965e24e2c","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":78,"snapshot_sha256":"e64df29f636a791e40ab5a6512eacf725807cb3017a7e7b37ab83bf67267770a","internal_anchors":1},"formal_canon":{"evidence_count":2,"snapshot_sha256":"0691e8785a0e1f57082d4837f1aad30c6038e8a61cf7eb396037b891d6767c86"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}