{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:AAY3BCOV7UMIHK3DF4XV7RVDKG","short_pith_number":"pith:AAY3BCOV","schema_version":"1.0","canonical_sha256":"0031b089d5fd1883ab632f2f5fc6a35194bb9b3810cd6975afa8a5b328fa9372","source":{"kind":"arxiv","id":"2410.04524","version":2},"attestation_state":"computed","paper":{"title":"Toward Secure Tuning: Mitigating Security Risks from Instruction Fine-Tuning","license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Bing Qin, Danyang Zhao, Fenglei Fan, Jiawei Cao, Ming Ma, Sendong Zhao, Shuren Qi, Ting Liu, Yanrui Du","submitted_at":"2024-10-06T15:34:04Z","abstract_excerpt":"Instruction fine-tuning has emerged as a critical technique for customizing Large Language Models (LLMs) to specific applications. However, recent studies have highlighted significant security vulnerabilities in fine-tuned LLMs. Existing defense efforts focus more on pre-training and post-training methods, yet there remains underexplored in in-training methods. To fill this gap, we introduce a novel secure-tuning strategy called SWAT. By analyzing how module-level parameters (e.g. Q/K/V/O) affect the security feature space drift, we identify a robust subset of modules, termed Mods_Rob. Our SWA"},"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":"2410.04524","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","primary_cat":"cs.CL","submitted_at":"2024-10-06T15:34:04Z","cross_cats_sorted":[],"title_canon_sha256":"3853a9d73ce09ec9c126298b887316f715de998699cc5b4803c058ed017d4fd6","abstract_canon_sha256":"e8231812a2f74367422d707d85f583543a8aaa0e532213eddc97ab5bba0fafae"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:15:19.718448Z","signature_b64":"Yh056Lec20PIzltA9Fu6GitoIoLMwnRnG1S8u6S1ZLpsTQ+liDNgoLoV2gaGZRisgjFe2YaN6GyHDxW7dfguBQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"0031b089d5fd1883ab632f2f5fc6a35194bb9b3810cd6975afa8a5b328fa9372","last_reissued_at":"2026-07-05T10:15:19.717950Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:15:19.717950Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Toward Secure Tuning: Mitigating Security Risks from Instruction Fine-Tuning","license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Bing Qin, Danyang Zhao, Fenglei Fan, Jiawei Cao, Ming Ma, Sendong Zhao, Shuren Qi, Ting Liu, Yanrui Du","submitted_at":"2024-10-06T15:34:04Z","abstract_excerpt":"Instruction fine-tuning has emerged as a critical technique for customizing Large Language Models (LLMs) to specific applications. However, recent studies have highlighted significant security vulnerabilities in fine-tuned LLMs. Existing defense efforts focus more on pre-training and post-training methods, yet there remains underexplored in in-training methods. To fill this gap, we introduce a novel secure-tuning strategy called SWAT. By analyzing how module-level parameters (e.g. Q/K/V/O) affect the security feature space drift, we identify a robust subset of modules, termed Mods_Rob. Our SWA"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2410.04524","kind":"arxiv","version":2},"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/2410.04524/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":"2410.04524","created_at":"2026-07-05T10:15:19.718033+00:00"},{"alias_kind":"arxiv_version","alias_value":"2410.04524v2","created_at":"2026-07-05T10:15:19.718033+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2410.04524","created_at":"2026-07-05T10:15:19.718033+00:00"},{"alias_kind":"pith_short_12","alias_value":"AAY3BCOV7UMI","created_at":"2026-07-05T10:15:19.718033+00:00"},{"alias_kind":"pith_short_16","alias_value":"AAY3BCOV7UMIHK3D","created_at":"2026-07-05T10:15:19.718033+00:00"},{"alias_kind":"pith_short_8","alias_value":"AAY3BCOV","created_at":"2026-07-05T10:15:19.718033+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":3,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.02530","citing_title":"SafeSteer: Localized On-Policy Distillation for Efficient Safety Alignment","ref_index":32,"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":39,"is_internal_anchor":false},{"citing_arxiv_id":"2601.14004","citing_title":"Locate, Steer, and Improve: A Practical Survey of Actionable Mechanistic Interpretability in Large Language Models","ref_index":62,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/AAY3BCOV7UMIHK3DF4XV7RVDKG","json":"https://pith.science/pith/AAY3BCOV7UMIHK3DF4XV7RVDKG.json","graph_json":"https://pith.science/api/pith-number/AAY3BCOV7UMIHK3DF4XV7RVDKG/graph.json","events_json":"https://pith.science/api/pith-number/AAY3BCOV7UMIHK3DF4XV7RVDKG/events.json","paper":"https://pith.science/paper/AAY3BCOV"},"agent_actions":{"view_html":"https://pith.science/pith/AAY3BCOV7UMIHK3DF4XV7RVDKG","download_json":"https://pith.science/pith/AAY3BCOV7UMIHK3DF4XV7RVDKG.json","view_paper":"https://pith.science/paper/AAY3BCOV","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2410.04524&json=true","fetch_graph":"https://pith.science/api/pith-number/AAY3BCOV7UMIHK3DF4XV7RVDKG/graph.json","fetch_events":"https://pith.science/api/pith-number/AAY3BCOV7UMIHK3DF4XV7RVDKG/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/AAY3BCOV7UMIHK3DF4XV7RVDKG/action/timestamp_anchor","attest_storage":"https://pith.science/pith/AAY3BCOV7UMIHK3DF4XV7RVDKG/action/storage_attestation","attest_author":"https://pith.science/pith/AAY3BCOV7UMIHK3DF4XV7RVDKG/action/author_attestation","sign_citation":"https://pith.science/pith/AAY3BCOV7UMIHK3DF4XV7RVDKG/action/citation_signature","submit_replication":"https://pith.science/pith/AAY3BCOV7UMIHK3DF4XV7RVDKG/action/replication_record"}},"created_at":"2026-07-05T10:15:19.718033+00:00","updated_at":"2026-07-05T10:15:19.718033+00:00"}