{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:UFKPMZDFOSF65UH5BUBH2TVIVD","short_pith_number":"pith:UFKPMZDF","schema_version":"1.0","canonical_sha256":"a154f66465748beed0fd0d027d4ea8a8ca6f58295f90f5628a0fbb309a6594b1","source":{"kind":"arxiv","id":"2401.04343","version":3},"attestation_state":"computed","paper":{"title":"Private Fine-tuning of Large Language Models with Zeroth-order Optimization","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"","cross_cats":["cs.CL","cs.CR"],"primary_cat":"cs.LG","authors_text":"Ashwinee Panda, Milad Nasr, Prateek Mittal, Saeed Mahloujifar, Xinyu Tang","submitted_at":"2024-01-09T03:53:59Z","abstract_excerpt":"Differentially private stochastic gradient descent (DP-SGD) allows models to be trained in a privacy-preserving manner, but has proven difficult to scale to the era of foundation models. We introduce DP-ZO, a private fine-tuning framework for large language models by privatizing zeroth order optimization methods. A key insight into the design of our method is that the direction of the gradient in the zeroth-order optimization we use is random and the only information from training data is the step size, i.e., a scalar. Therefore, we only need to privatize the scalar step size, which is memory-"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":false,"formal_links_present":false},"canonical_record":{"source":{"id":"2401.04343","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by-sa/4.0/","primary_cat":"cs.LG","submitted_at":"2024-01-09T03:53:59Z","cross_cats_sorted":["cs.CL","cs.CR"],"title_canon_sha256":"73c4c0580fe97aba53eaa74adc0432d708c5782f7f8a2b03750351548f5c91e1","abstract_canon_sha256":"2f292722e245360003514449c750f825c2aa4dc6c2fc7d3846baebbfd510d79b"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:07:38.976698Z","signature_b64":"CdWQD8BNL8XoH/l3jbDBn6C4mDI7kduk5C7/SYfmFnBgXRn1HFn4qqiiuK23neVgoeRhLthQeLX3jCKYsJu8CA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"a154f66465748beed0fd0d027d4ea8a8ca6f58295f90f5628a0fbb309a6594b1","last_reissued_at":"2026-07-05T10:07:38.975962Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:07:38.975962Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Private Fine-tuning of Large Language Models with Zeroth-order Optimization","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"","cross_cats":["cs.CL","cs.CR"],"primary_cat":"cs.LG","authors_text":"Ashwinee Panda, Milad Nasr, Prateek Mittal, Saeed Mahloujifar, Xinyu Tang","submitted_at":"2024-01-09T03:53:59Z","abstract_excerpt":"Differentially private stochastic gradient descent (DP-SGD) allows models to be trained in a privacy-preserving manner, but has proven difficult to scale to the era of foundation models. We introduce DP-ZO, a private fine-tuning framework for large language models by privatizing zeroth order optimization methods. A key insight into the design of our method is that the direction of the gradient in the zeroth-order optimization we use is random and the only information from training data is the step size, i.e., a scalar. Therefore, we only need to privatize the scalar step size, which is memory-"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2401.04343","kind":"arxiv","version":3},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2401.04343/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"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"},"aliases":[{"alias_kind":"arxiv","alias_value":"2401.04343","created_at":"2026-07-05T10:07:38.976013+00:00"},{"alias_kind":"arxiv_version","alias_value":"2401.04343v3","created_at":"2026-07-05T10:07:38.976013+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2401.04343","created_at":"2026-07-05T10:07:38.976013+00:00"},{"alias_kind":"pith_short_12","alias_value":"UFKPMZDFOSF6","created_at":"2026-07-05T10:07:38.976013+00:00"},{"alias_kind":"pith_short_16","alias_value":"UFKPMZDFOSF65UH5","created_at":"2026-07-05T10:07:38.976013+00:00"},{"alias_kind":"pith_short_8","alias_value":"UFKPMZDF","created_at":"2026-07-05T10:07:38.976013+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":6,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.09125","citing_title":"Unveiling Privacy Risks in Multi-modal Large Language Models: Task-specific Vulnerabilities and Mitigation Challenges","ref_index":72,"is_internal_anchor":false},{"citing_arxiv_id":"2606.31208","citing_title":"Probing Memorization of Tabular In-Context Learning","ref_index":52,"is_internal_anchor":false},{"citing_arxiv_id":"2605.24879","citing_title":"Efficient DP-SGD for LLMs with Randomized Clipping","ref_index":46,"is_internal_anchor":false},{"citing_arxiv_id":"2506.15588","citing_title":"Memory-Efficient Differentially Private Training with Gradient Random Projection","ref_index":17,"is_internal_anchor":false},{"citing_arxiv_id":"2506.00158","citing_title":"Privacy Amplification in Differentially Private Zeroth-Order Optimization with Hidden States","ref_index":23,"is_internal_anchor":false},{"citing_arxiv_id":"2601.10237","citing_title":"Fundamental Limitations of Favorable Privacy-Utility Guarantees for DP-SGD","ref_index":59,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/UFKPMZDFOSF65UH5BUBH2TVIVD","json":"https://pith.science/pith/UFKPMZDFOSF65UH5BUBH2TVIVD.json","graph_json":"https://pith.science/api/pith-number/UFKPMZDFOSF65UH5BUBH2TVIVD/graph.json","events_json":"https://pith.science/api/pith-number/UFKPMZDFOSF65UH5BUBH2TVIVD/events.json","paper":"https://pith.science/paper/UFKPMZDF"},"agent_actions":{"view_html":"https://pith.science/pith/UFKPMZDFOSF65UH5BUBH2TVIVD","download_json":"https://pith.science/pith/UFKPMZDFOSF65UH5BUBH2TVIVD.json","view_paper":"https://pith.science/paper/UFKPMZDF","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2401.04343&json=true","fetch_graph":"https://pith.science/api/pith-number/UFKPMZDFOSF65UH5BUBH2TVIVD/graph.json","fetch_events":"https://pith.science/api/pith-number/UFKPMZDFOSF65UH5BUBH2TVIVD/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/UFKPMZDFOSF65UH5BUBH2TVIVD/action/timestamp_anchor","attest_storage":"https://pith.science/pith/UFKPMZDFOSF65UH5BUBH2TVIVD/action/storage_attestation","attest_author":"https://pith.science/pith/UFKPMZDFOSF65UH5BUBH2TVIVD/action/author_attestation","sign_citation":"https://pith.science/pith/UFKPMZDFOSF65UH5BUBH2TVIVD/action/citation_signature","submit_replication":"https://pith.science/pith/UFKPMZDFOSF65UH5BUBH2TVIVD/action/replication_record"}},"created_at":"2026-07-05T10:07:38.976013+00:00","updated_at":"2026-07-05T10:07:38.976013+00:00"}