{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:PRY25LEOQNUKHXN7U73XQ66OFY","short_pith_number":"pith:PRY25LEO","schema_version":"1.0","canonical_sha256":"7c71aeac8e8368a3ddbfa7f7787bce2e07ab9c664d1dafbef358901e2d5f5b1e","source":{"kind":"arxiv","id":"2406.07778","version":2},"attestation_state":"computed","paper":{"title":"A Study of Backdoors in Instruction Fine-tuned Language Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.CL","cs.LG"],"primary_cat":"cs.CR","authors_text":"David J. Miller, George Kesidis, Jayaram Raghuram","submitted_at":"2024-06-12T00:01:32Z","abstract_excerpt":"Backdoor data poisoning, inserted within instruction examples used to fine-tune a foundation Large Language Model (LLM) for downstream tasks (\\textit{e.g.,} sentiment prediction), is a serious security concern due to the evasive nature of such attacks. The poisoning is usually in the form of a (seemingly innocuous) trigger word or phrase inserted into a very small fraction of the fine-tuning samples from a target class. Such backdoor attacks can: alter response sentiment, violate censorship, over-refuse (invoke censorship for legitimate queries), inject false content, or trigger nonsense respo"},"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":"2406.07778","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CR","submitted_at":"2024-06-12T00:01:32Z","cross_cats_sorted":["cs.AI","cs.CL","cs.LG"],"title_canon_sha256":"144f74023595510a4507c3ba953d173b618ba375fb5cb3417cf3aae14689ae53","abstract_canon_sha256":"ff80b6ca95cb6e54a7cbad33f0283b6b52bebc468abcd37d81930b50386a9243"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T08:57:57.157012Z","signature_b64":"ZKJNoZnCo5+kgT6d73v5Ef+4ZVdrjfHDXerr/dT1f5r9rYfyUB9JxQhqVDOeEvjk3gQulwi76Ef/NDSlEmYSBw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"7c71aeac8e8368a3ddbfa7f7787bce2e07ab9c664d1dafbef358901e2d5f5b1e","last_reissued_at":"2026-07-05T08:57:57.156581Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T08:57:57.156581Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"A Study of Backdoors in Instruction Fine-tuned Language Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.CL","cs.LG"],"primary_cat":"cs.CR","authors_text":"David J. Miller, George Kesidis, Jayaram Raghuram","submitted_at":"2024-06-12T00:01:32Z","abstract_excerpt":"Backdoor data poisoning, inserted within instruction examples used to fine-tune a foundation Large Language Model (LLM) for downstream tasks (\\textit{e.g.,} sentiment prediction), is a serious security concern due to the evasive nature of such attacks. The poisoning is usually in the form of a (seemingly innocuous) trigger word or phrase inserted into a very small fraction of the fine-tuning samples from a target class. Such backdoor attacks can: alter response sentiment, violate censorship, over-refuse (invoke censorship for legitimate queries), inject false content, or trigger nonsense respo"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2406.07778","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/2406.07778/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":"2406.07778","created_at":"2026-07-05T08:57:57.156650+00:00"},{"alias_kind":"arxiv_version","alias_value":"2406.07778v2","created_at":"2026-07-05T08:57:57.156650+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2406.07778","created_at":"2026-07-05T08:57:57.156650+00:00"},{"alias_kind":"pith_short_12","alias_value":"PRY25LEOQNUK","created_at":"2026-07-05T08:57:57.156650+00:00"},{"alias_kind":"pith_short_16","alias_value":"PRY25LEOQNUKHXN7","created_at":"2026-07-05T08:57:57.156650+00:00"},{"alias_kind":"pith_short_8","alias_value":"PRY25LEO","created_at":"2026-07-05T08:57:57.156650+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":5,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2607.05300","citing_title":"Learning Only What Valid Adapters Can Express: Subspace-Constrained Adaptation Against Fine-Tuning Poisoning","ref_index":9,"is_internal_anchor":true},{"citing_arxiv_id":"2605.19262","citing_title":"Backdooring Masked Diffusion Language Models","ref_index":37,"is_internal_anchor":false},{"citing_arxiv_id":"2605.19262","citing_title":"Backdooring Masked Diffusion Language Models","ref_index":37,"is_internal_anchor":false},{"citing_arxiv_id":"2604.09378","citing_title":"BadSkill: Backdoor Attacks on Agent Skills via Model-in-Skill Poisoning","ref_index":19,"is_internal_anchor":false},{"citing_arxiv_id":"2604.04738","citing_title":"Fine-Tuning Integrity for Modern Neural Networks: Structured Drift Proofs via Norm, Rank, and Sparsity Certificates","ref_index":33,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/PRY25LEOQNUKHXN7U73XQ66OFY","json":"https://pith.science/pith/PRY25LEOQNUKHXN7U73XQ66OFY.json","graph_json":"https://pith.science/api/pith-number/PRY25LEOQNUKHXN7U73XQ66OFY/graph.json","events_json":"https://pith.science/api/pith-number/PRY25LEOQNUKHXN7U73XQ66OFY/events.json","paper":"https://pith.science/paper/PRY25LEO"},"agent_actions":{"view_html":"https://pith.science/pith/PRY25LEOQNUKHXN7U73XQ66OFY","download_json":"https://pith.science/pith/PRY25LEOQNUKHXN7U73XQ66OFY.json","view_paper":"https://pith.science/paper/PRY25LEO","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2406.07778&json=true","fetch_graph":"https://pith.science/api/pith-number/PRY25LEOQNUKHXN7U73XQ66OFY/graph.json","fetch_events":"https://pith.science/api/pith-number/PRY25LEOQNUKHXN7U73XQ66OFY/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/PRY25LEOQNUKHXN7U73XQ66OFY/action/timestamp_anchor","attest_storage":"https://pith.science/pith/PRY25LEOQNUKHXN7U73XQ66OFY/action/storage_attestation","attest_author":"https://pith.science/pith/PRY25LEOQNUKHXN7U73XQ66OFY/action/author_attestation","sign_citation":"https://pith.science/pith/PRY25LEOQNUKHXN7U73XQ66OFY/action/citation_signature","submit_replication":"https://pith.science/pith/PRY25LEOQNUKHXN7U73XQ66OFY/action/replication_record"}},"created_at":"2026-07-05T08:57:57.156650+00:00","updated_at":"2026-07-05T08:57:57.156650+00:00"}