{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:BDOGV5CB3URXX4RP2AWLEJ7TIH","short_pith_number":"pith:BDOGV5CB","schema_version":"1.0","canonical_sha256":"08dc6af441dd237bf22fd02cb227f341c59d0094018d548ac729adee0cabe546","source":{"kind":"arxiv","id":"2408.09093","version":3},"attestation_state":"computed","paper":{"title":"BaThe: Defense against the Jailbreak Attack in Multimodal Large Language Models by Treating Harmful Instruction as Backdoor Trigger","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CR","authors_text":"Bryan Hooi, Haoran Li, Yangqiu Song, Yirui Zhang, Yulin Chen, Zihao Zheng","submitted_at":"2024-08-17T04:43:26Z","abstract_excerpt":"Multimodal Large Language Models (MLLMs) have showcased impressive performance in a variety of multimodal tasks. On the other hand, the integration of additional image modality may allow the malicious users to inject harmful content inside the images for jailbreaking. Unlike text-based LLMs, where adversaries need to select discrete tokens to conceal their malicious intent using specific algorithms, the continuous nature of image signals provides a direct opportunity for adversaries to inject harmful intentions. In this work, we propose $\\textbf{BaThe}$ ($\\textbf{Ba}$ckdoor $\\textbf{T}$rigger "},"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":"2408.09093","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CR","submitted_at":"2024-08-17T04:43:26Z","cross_cats_sorted":[],"title_canon_sha256":"650da74a1335e2816e0cc4bf4a2e33328f0d040d517d26fb19c63e3c127e90cf","abstract_canon_sha256":"7863cb665a79370cfaf74f9f27662408fbcd9c1063da7587ea66302214fc8c47"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:52:13.490254Z","signature_b64":"8JB+wmh5o76FyzsptjuCtXVzyCliUan3RayPYmDj21eqC02WQcLd9J1S55rJc344SfamPNCRG3NIqrZKgbspCg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"08dc6af441dd237bf22fd02cb227f341c59d0094018d548ac729adee0cabe546","last_reissued_at":"2026-07-05T10:52:13.489761Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:52:13.489761Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"BaThe: Defense against the Jailbreak Attack in Multimodal Large Language Models by Treating Harmful Instruction as Backdoor Trigger","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CR","authors_text":"Bryan Hooi, Haoran Li, Yangqiu Song, Yirui Zhang, Yulin Chen, Zihao Zheng","submitted_at":"2024-08-17T04:43:26Z","abstract_excerpt":"Multimodal Large Language Models (MLLMs) have showcased impressive performance in a variety of multimodal tasks. On the other hand, the integration of additional image modality may allow the malicious users to inject harmful content inside the images for jailbreaking. Unlike text-based LLMs, where adversaries need to select discrete tokens to conceal their malicious intent using specific algorithms, the continuous nature of image signals provides a direct opportunity for adversaries to inject harmful intentions. In this work, we propose $\\textbf{BaThe}$ ($\\textbf{Ba}$ckdoor $\\textbf{T}$rigger "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2408.09093","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/2408.09093/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":"2408.09093","created_at":"2026-07-05T10:52:13.489817+00:00"},{"alias_kind":"arxiv_version","alias_value":"2408.09093v3","created_at":"2026-07-05T10:52:13.489817+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2408.09093","created_at":"2026-07-05T10:52:13.489817+00:00"},{"alias_kind":"pith_short_12","alias_value":"BDOGV5CB3URX","created_at":"2026-07-05T10:52:13.489817+00:00"},{"alias_kind":"pith_short_16","alias_value":"BDOGV5CB3URXX4RP","created_at":"2026-07-05T10:52:13.489817+00:00"},{"alias_kind":"pith_short_8","alias_value":"BDOGV5CB","created_at":"2026-07-05T10:52:13.489817+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2506.05401","citing_title":"Robust Anti-Backdoor Instruction Tuning in LVLMs","ref_index":24,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/BDOGV5CB3URXX4RP2AWLEJ7TIH","json":"https://pith.science/pith/BDOGV5CB3URXX4RP2AWLEJ7TIH.json","graph_json":"https://pith.science/api/pith-number/BDOGV5CB3URXX4RP2AWLEJ7TIH/graph.json","events_json":"https://pith.science/api/pith-number/BDOGV5CB3URXX4RP2AWLEJ7TIH/events.json","paper":"https://pith.science/paper/BDOGV5CB"},"agent_actions":{"view_html":"https://pith.science/pith/BDOGV5CB3URXX4RP2AWLEJ7TIH","download_json":"https://pith.science/pith/BDOGV5CB3URXX4RP2AWLEJ7TIH.json","view_paper":"https://pith.science/paper/BDOGV5CB","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2408.09093&json=true","fetch_graph":"https://pith.science/api/pith-number/BDOGV5CB3URXX4RP2AWLEJ7TIH/graph.json","fetch_events":"https://pith.science/api/pith-number/BDOGV5CB3URXX4RP2AWLEJ7TIH/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/BDOGV5CB3URXX4RP2AWLEJ7TIH/action/timestamp_anchor","attest_storage":"https://pith.science/pith/BDOGV5CB3URXX4RP2AWLEJ7TIH/action/storage_attestation","attest_author":"https://pith.science/pith/BDOGV5CB3URXX4RP2AWLEJ7TIH/action/author_attestation","sign_citation":"https://pith.science/pith/BDOGV5CB3URXX4RP2AWLEJ7TIH/action/citation_signature","submit_replication":"https://pith.science/pith/BDOGV5CB3URXX4RP2AWLEJ7TIH/action/replication_record"}},"created_at":"2026-07-05T10:52:13.489817+00:00","updated_at":"2026-07-05T10:52:13.489817+00:00"}