{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:2XXBQPRUIS2QI7SXUBYAPU3NT5","short_pith_number":"pith:2XXBQPRU","schema_version":"1.0","canonical_sha256":"d5ee183e3444b5047e57a07007d36d9f587af13cbfd946c71725705bd1c43248","source":{"kind":"arxiv","id":"2405.14577","version":4},"attestation_state":"computed","paper":{"title":"Representation Noising: A Defence Mechanism Against Harmful Finetuning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Carsten Maple, David Atanasov, Domenic Rosati, Frank Rudzicz, Hassan Sajjad, Jan Wehner, Kai Williams, {\\L}ukasz Bartoszcze, Robie Gonzales, Subhabrata Majumdar","submitted_at":"2024-05-23T13:51:55Z","abstract_excerpt":"Releasing open-source large language models (LLMs) presents a dual-use risk since bad actors can easily fine-tune these models for harmful purposes. Even without the open release of weights, weight stealing and fine-tuning APIs make closed models vulnerable to harmful fine-tuning attacks (HFAs). While safety measures like preventing jailbreaks and improving safety guardrails are important, such measures can easily be reversed through fine-tuning. In this work, we propose Representation Noising (RepNoise), a defence mechanism that operates even when attackers have access to the weights. RepNois"},"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":"2405.14577","kind":"arxiv","version":4},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2024-05-23T13:51:55Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"b80411dddc6468e80a7a0a898f8d49761747df37fdb1295b55b23baf3dcb81cf","abstract_canon_sha256":"d41c0b506c361ad90ceb84580c874b6e582e8f994d0ff9e67a37722cb87fc837"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T09:28:49.966605Z","signature_b64":"vN21S2r61opE8sqZxk2QzhUsQakYbZGbc+5xprOAs4FuXpu/2blhA+b/jsPZejqQNxDmuwGn6cI5RLKTSeG7AA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"d5ee183e3444b5047e57a07007d36d9f587af13cbfd946c71725705bd1c43248","last_reissued_at":"2026-07-05T09:28:49.966104Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T09:28:49.966104Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Representation Noising: A Defence Mechanism Against Harmful Finetuning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Carsten Maple, David Atanasov, Domenic Rosati, Frank Rudzicz, Hassan Sajjad, Jan Wehner, Kai Williams, {\\L}ukasz Bartoszcze, Robie Gonzales, Subhabrata Majumdar","submitted_at":"2024-05-23T13:51:55Z","abstract_excerpt":"Releasing open-source large language models (LLMs) presents a dual-use risk since bad actors can easily fine-tune these models for harmful purposes. Even without the open release of weights, weight stealing and fine-tuning APIs make closed models vulnerable to harmful fine-tuning attacks (HFAs). While safety measures like preventing jailbreaks and improving safety guardrails are important, such measures can easily be reversed through fine-tuning. In this work, we propose Representation Noising (RepNoise), a defence mechanism that operates even when attackers have access to the weights. RepNois"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2405.14577","kind":"arxiv","version":4},"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/2405.14577/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":"2405.14577","created_at":"2026-07-05T09:28:49.966162+00:00"},{"alias_kind":"arxiv_version","alias_value":"2405.14577v4","created_at":"2026-07-05T09:28:49.966162+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2405.14577","created_at":"2026-07-05T09:28:49.966162+00:00"},{"alias_kind":"pith_short_12","alias_value":"2XXBQPRUIS2Q","created_at":"2026-07-05T09:28:49.966162+00:00"},{"alias_kind":"pith_short_16","alias_value":"2XXBQPRUIS2QI7SX","created_at":"2026-07-05T09:28:49.966162+00:00"},{"alias_kind":"pith_short_8","alias_value":"2XXBQPRU","created_at":"2026-07-05T09:28:49.966162+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":5,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.23075","citing_title":"Safety in Self-Evolving LLM Agent Systems: Threats, Amplification, and Case Studies","ref_index":61,"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":128,"is_internal_anchor":false},{"citing_arxiv_id":"2602.17546","citing_title":"Learning to Stay Safe: Adaptive Regularization Against Safety Degradation during Fine-Tuning","ref_index":7,"is_internal_anchor":false},{"citing_arxiv_id":"2604.09688","citing_title":"Immunizing 3D Gaussian Generative Models Against Unauthorized Fine-Tuning via Attribute-Space Traps","ref_index":33,"is_internal_anchor":false},{"citing_arxiv_id":"2604.17215","citing_title":"Continual Safety Alignment via Gradient-Based Sample Selection","ref_index":3,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/2XXBQPRUIS2QI7SXUBYAPU3NT5","json":"https://pith.science/pith/2XXBQPRUIS2QI7SXUBYAPU3NT5.json","graph_json":"https://pith.science/api/pith-number/2XXBQPRUIS2QI7SXUBYAPU3NT5/graph.json","events_json":"https://pith.science/api/pith-number/2XXBQPRUIS2QI7SXUBYAPU3NT5/events.json","paper":"https://pith.science/paper/2XXBQPRU"},"agent_actions":{"view_html":"https://pith.science/pith/2XXBQPRUIS2QI7SXUBYAPU3NT5","download_json":"https://pith.science/pith/2XXBQPRUIS2QI7SXUBYAPU3NT5.json","view_paper":"https://pith.science/paper/2XXBQPRU","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2405.14577&json=true","fetch_graph":"https://pith.science/api/pith-number/2XXBQPRUIS2QI7SXUBYAPU3NT5/graph.json","fetch_events":"https://pith.science/api/pith-number/2XXBQPRUIS2QI7SXUBYAPU3NT5/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/2XXBQPRUIS2QI7SXUBYAPU3NT5/action/timestamp_anchor","attest_storage":"https://pith.science/pith/2XXBQPRUIS2QI7SXUBYAPU3NT5/action/storage_attestation","attest_author":"https://pith.science/pith/2XXBQPRUIS2QI7SXUBYAPU3NT5/action/author_attestation","sign_citation":"https://pith.science/pith/2XXBQPRUIS2QI7SXUBYAPU3NT5/action/citation_signature","submit_replication":"https://pith.science/pith/2XXBQPRUIS2QI7SXUBYAPU3NT5/action/replication_record"}},"created_at":"2026-07-05T09:28:49.966162+00:00","updated_at":"2026-07-05T09:28:49.966162+00:00"}