{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:W4DYHNODJBYTR3EITNJARKZZOT","short_pith_number":"pith:W4DYHNOD","schema_version":"1.0","canonical_sha256":"b70783b5c3487138ec889b5208ab3974e7484164ff5b9739771998532cdee434","source":{"kind":"arxiv","id":"2403.15638","version":3},"attestation_state":"computed","paper":{"title":"Differentially Private Next-Token Prediction of Large Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL","cs.LG"],"primary_cat":"cs.CR","authors_text":"James Flemings, Meisam Razaviyayn, Murali Annavaram","submitted_at":"2024-03-22T22:27:44Z","abstract_excerpt":"Ensuring the privacy of Large Language Models (LLMs) is becoming increasingly important. The most widely adopted technique to accomplish this is DP-SGD, which trains a model to guarantee Differential Privacy (DP). However, DP-SGD overestimates an adversary's capabilities in having white box access to the model and, as a result, causes longer training times and larger memory usage than SGD. On the other hand, commercial LLM deployments are predominantly cloud-based; hence, adversarial access to LLMs is black-box. Motivated by these observations, we present Private Mixing of Ensemble Distributio"},"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":"2403.15638","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CR","submitted_at":"2024-03-22T22:27:44Z","cross_cats_sorted":["cs.CL","cs.LG"],"title_canon_sha256":"69924b06081214d6773b3c479acf5d32975e8e5fec197cc85737d454f65628b9","abstract_canon_sha256":"bed35b8e4e60d7390c8641c8668e3d0712dbd6451c7929442e8ee5f8e4060f65"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T08:12:49.151262Z","signature_b64":"53glUbRzG1WfHBbTHPbwE7zFHpTm9CXSZuxGDtCRaFwpuQqoWQEhGF38gUjy7RlUee+1O97QjM2tglnzjudyBw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"b70783b5c3487138ec889b5208ab3974e7484164ff5b9739771998532cdee434","last_reissued_at":"2026-07-05T08:12:49.150825Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T08:12:49.150825Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Differentially Private Next-Token Prediction of Large Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL","cs.LG"],"primary_cat":"cs.CR","authors_text":"James Flemings, Meisam Razaviyayn, Murali Annavaram","submitted_at":"2024-03-22T22:27:44Z","abstract_excerpt":"Ensuring the privacy of Large Language Models (LLMs) is becoming increasingly important. The most widely adopted technique to accomplish this is DP-SGD, which trains a model to guarantee Differential Privacy (DP). However, DP-SGD overestimates an adversary's capabilities in having white box access to the model and, as a result, causes longer training times and larger memory usage than SGD. On the other hand, commercial LLM deployments are predominantly cloud-based; hence, adversarial access to LLMs is black-box. Motivated by these observations, we present Private Mixing of Ensemble Distributio"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2403.15638","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/2403.15638/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":"2403.15638","created_at":"2026-07-05T08:12:49.150879+00:00"},{"alias_kind":"arxiv_version","alias_value":"2403.15638v3","created_at":"2026-07-05T08:12:49.150879+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2403.15638","created_at":"2026-07-05T08:12:49.150879+00:00"},{"alias_kind":"pith_short_12","alias_value":"W4DYHNODJBYT","created_at":"2026-07-05T08:12:49.150879+00:00"},{"alias_kind":"pith_short_16","alias_value":"W4DYHNODJBYTR3EI","created_at":"2026-07-05T08:12:49.150879+00:00"},{"alias_kind":"pith_short_8","alias_value":"W4DYHNOD","created_at":"2026-07-05T08:12:49.150879+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2509.12958","citing_title":"Forget What's Sensitive, Remember What Matters: Token-Level Differential Privacy in Memory Sculpting for Continual Learning","ref_index":13,"is_internal_anchor":false},{"citing_arxiv_id":"2507.02974","citing_title":"InvisibleInk: High-Utility and Low-Cost Text Generation with Differential Privacy","ref_index":49,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/W4DYHNODJBYTR3EITNJARKZZOT","json":"https://pith.science/pith/W4DYHNODJBYTR3EITNJARKZZOT.json","graph_json":"https://pith.science/api/pith-number/W4DYHNODJBYTR3EITNJARKZZOT/graph.json","events_json":"https://pith.science/api/pith-number/W4DYHNODJBYTR3EITNJARKZZOT/events.json","paper":"https://pith.science/paper/W4DYHNOD"},"agent_actions":{"view_html":"https://pith.science/pith/W4DYHNODJBYTR3EITNJARKZZOT","download_json":"https://pith.science/pith/W4DYHNODJBYTR3EITNJARKZZOT.json","view_paper":"https://pith.science/paper/W4DYHNOD","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2403.15638&json=true","fetch_graph":"https://pith.science/api/pith-number/W4DYHNODJBYTR3EITNJARKZZOT/graph.json","fetch_events":"https://pith.science/api/pith-number/W4DYHNODJBYTR3EITNJARKZZOT/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/W4DYHNODJBYTR3EITNJARKZZOT/action/timestamp_anchor","attest_storage":"https://pith.science/pith/W4DYHNODJBYTR3EITNJARKZZOT/action/storage_attestation","attest_author":"https://pith.science/pith/W4DYHNODJBYTR3EITNJARKZZOT/action/author_attestation","sign_citation":"https://pith.science/pith/W4DYHNODJBYTR3EITNJARKZZOT/action/citation_signature","submit_replication":"https://pith.science/pith/W4DYHNODJBYTR3EITNJARKZZOT/action/replication_record"}},"created_at":"2026-07-05T08:12:49.150879+00:00","updated_at":"2026-07-05T08:12:49.150879+00:00"}