{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:VS74OPA6Z6M2SXOTLRY47AGCSS","short_pith_number":"pith:VS74OPA6","schema_version":"1.0","canonical_sha256":"acbfc73c1ecf99a95dd35c71cf80c29485f82367361cb88d71be3ae2a04f7f1e","source":{"kind":"arxiv","id":"2408.09600","version":3},"attestation_state":"computed","paper":{"title":"Antidote: Post-fine-tuning Safety Alignment for Large Language Models against Harmful Fine-tuning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CR"],"primary_cat":"cs.AI","authors_text":"Gautam Bhattacharya, Josh Kimball, Ling Liu, Pratik Joshi, Tiansheng Huang","submitted_at":"2024-08-18T21:45:03Z","abstract_excerpt":"Safety aligned Large Language Models (LLMs) are vulnerable to harmful fine-tuning attacks -- a few harmful data mixed in the fine-tuning dataset can break the LLMs's safety alignment. While several defenses have been proposed, our evaluation shows that existing defenses fail \\textit{when some specific training hyper-parameters are chosen} -- a large learning rate or a large number of training epochs in the fine-tuning stage can easily invalidate the defense. To this end, we propose Antidote, a post-fine-tuning stage solution, which remains \\textbf{\\textit{agnostic to the training hyper-paramet"},"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":"2408.09600","kind":"arxiv","version":3},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.AI","submitted_at":"2024-08-18T21:45:03Z","cross_cats_sorted":["cs.CR"],"title_canon_sha256":"cc677f136fd792f9f37796162259c7ed5ea0044facd8b5a835f8bc934b2d79d6","abstract_canon_sha256":"8cafc6f2c643ef58b7d5ca99915224c082e5a4ec96f8e0ed4c4dc412c313d93e"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T12:05:09.792311Z","signature_b64":"6vgFt8tdcJl4v+A+T8EiImgtN91CGnoOVtYzYw2suLB4YsF/qiJyvPjIpDo/44zPEreOpaAn5TyUtwEZ1x6LDg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"acbfc73c1ecf99a95dd35c71cf80c29485f82367361cb88d71be3ae2a04f7f1e","last_reissued_at":"2026-07-05T12:05:09.791778Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T12:05:09.791778Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Antidote: Post-fine-tuning Safety Alignment for Large Language Models against Harmful Fine-tuning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.CR"],"primary_cat":"cs.AI","authors_text":"Gautam Bhattacharya, Josh Kimball, Ling Liu, Pratik Joshi, Tiansheng Huang","submitted_at":"2024-08-18T21:45:03Z","abstract_excerpt":"Safety aligned Large Language Models (LLMs) are vulnerable to harmful fine-tuning attacks -- a few harmful data mixed in the fine-tuning dataset can break the LLMs's safety alignment. While several defenses have been proposed, our evaluation shows that existing defenses fail \\textit{when some specific training hyper-parameters are chosen} -- a large learning rate or a large number of training epochs in the fine-tuning stage can easily invalidate the defense. To this end, we propose Antidote, a post-fine-tuning stage solution, which remains \\textbf{\\textit{agnostic to the training hyper-paramet"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2408.09600","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/2408.09600/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":"2408.09600","created_at":"2026-07-05T12:05:09.791837+00:00"},{"alias_kind":"arxiv_version","alias_value":"2408.09600v3","created_at":"2026-07-05T12:05:09.791837+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2408.09600","created_at":"2026-07-05T12:05:09.791837+00:00"},{"alias_kind":"pith_short_12","alias_value":"VS74OPA6Z6M2","created_at":"2026-07-05T12:05:09.791837+00:00"},{"alias_kind":"pith_short_16","alias_value":"VS74OPA6Z6M2SXOT","created_at":"2026-07-05T12:05:09.791837+00:00"},{"alias_kind":"pith_short_8","alias_value":"VS74OPA6","created_at":"2026-07-05T12:05:09.791837+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":7,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.07970","citing_title":"Defending Against Malicious Finetuning by Scaling Train-time Adversarial Attacks","ref_index":32,"is_internal_anchor":false},{"citing_arxiv_id":"2606.06519","citing_title":"SafeGene: Reusable Adapters for Transferable Safety Alignment","ref_index":6,"is_internal_anchor":false},{"citing_arxiv_id":"2409.18169","citing_title":"Harmful Fine-tuning Attacks and Defenses for Large Language Models: A Survey","ref_index":64,"is_internal_anchor":false},{"citing_arxiv_id":"2502.05206","citing_title":"Safety at Scale: A Comprehensive Survey of Large Model and Agent Safety","ref_index":123,"is_internal_anchor":false},{"citing_arxiv_id":"2605.14194","citing_title":"GradShield: Alignment Preserving Finetuning","ref_index":24,"is_internal_anchor":false},{"citing_arxiv_id":"2604.24536","citing_title":"Generating Place-Based Compromises Between Two Points of View","ref_index":35,"is_internal_anchor":false},{"citing_arxiv_id":"2604.12384","citing_title":"Preventing Safety Drift in Large Language Models via Coupled Weight and Activation Constraints","ref_index":14,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/VS74OPA6Z6M2SXOTLRY47AGCSS","json":"https://pith.science/pith/VS74OPA6Z6M2SXOTLRY47AGCSS.json","graph_json":"https://pith.science/api/pith-number/VS74OPA6Z6M2SXOTLRY47AGCSS/graph.json","events_json":"https://pith.science/api/pith-number/VS74OPA6Z6M2SXOTLRY47AGCSS/events.json","paper":"https://pith.science/paper/VS74OPA6"},"agent_actions":{"view_html":"https://pith.science/pith/VS74OPA6Z6M2SXOTLRY47AGCSS","download_json":"https://pith.science/pith/VS74OPA6Z6M2SXOTLRY47AGCSS.json","view_paper":"https://pith.science/paper/VS74OPA6","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2408.09600&json=true","fetch_graph":"https://pith.science/api/pith-number/VS74OPA6Z6M2SXOTLRY47AGCSS/graph.json","fetch_events":"https://pith.science/api/pith-number/VS74OPA6Z6M2SXOTLRY47AGCSS/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/VS74OPA6Z6M2SXOTLRY47AGCSS/action/timestamp_anchor","attest_storage":"https://pith.science/pith/VS74OPA6Z6M2SXOTLRY47AGCSS/action/storage_attestation","attest_author":"https://pith.science/pith/VS74OPA6Z6M2SXOTLRY47AGCSS/action/author_attestation","sign_citation":"https://pith.science/pith/VS74OPA6Z6M2SXOTLRY47AGCSS/action/citation_signature","submit_replication":"https://pith.science/pith/VS74OPA6Z6M2SXOTLRY47AGCSS/action/replication_record"}},"created_at":"2026-07-05T12:05:09.791837+00:00","updated_at":"2026-07-05T12:05:09.791837+00:00"}