{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:7PZ65OX2Y3D4XZ7IKHJK5IVOA5","short_pith_number":"pith:7PZ65OX2","schema_version":"1.0","canonical_sha256":"fbf3eebafac6c7cbe7e851d2aea2ae074413e43b490a772e77d91a3f539294ef","source":{"kind":"arxiv","id":"2601.21692","version":2},"attestation_state":"computed","paper":{"title":"TCAP: Tri-Component Attention Profiling for Unsupervised Backdoor Detection in MLLM Fine-Tuning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Hao Fang, Mingzu Liu, Runmin Cong","submitted_at":"2026-01-29T13:26:29Z","abstract_excerpt":"Fine-Tuning-as-a-Service (FTaaS) facilitates the customization of Multimodal Large Language Models (MLLMs) but introduces critical backdoor risks via poisoned data. Existing defenses either rely on supervised signals or fail to generalize across diverse trigger types and modalities. In this work, we uncover a universal backdoor fingerprint-attention allocation divergence-where poisoned samples disrupt the balanced attention distribution across three functional components: system instructions, vision inputs, and user textual queries, regardless of trigger morphology. Motivated by this insight, "},"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":"2601.21692","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.AI","submitted_at":"2026-01-29T13:26:29Z","cross_cats_sorted":[],"title_canon_sha256":"641c96d888d4103c45370196b9387d5d54534bb7d3056ecee5f8565e745e22c9","abstract_canon_sha256":"ce0bdfae0a7a4340d3b4aa65271be4a810c636264c6fdd4180ac7cd7a4064f58"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-25T02:01:13.606739Z","signature_b64":"rml9/0h9sSPfDCEmCZaHI1j3/dhKqfxvkBIv9LYtmj65PDut1kkMko9vpq2yDBacH7dYJ9UQAvW1AJ1902/bBg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"fbf3eebafac6c7cbe7e851d2aea2ae074413e43b490a772e77d91a3f539294ef","last_reissued_at":"2026-05-25T02:01:13.606144Z","signature_status":"signed_v1","first_computed_at":"2026-05-25T02:01:13.606144Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"TCAP: Tri-Component Attention Profiling for Unsupervised Backdoor Detection in MLLM Fine-Tuning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Hao Fang, Mingzu Liu, Runmin Cong","submitted_at":"2026-01-29T13:26:29Z","abstract_excerpt":"Fine-Tuning-as-a-Service (FTaaS) facilitates the customization of Multimodal Large Language Models (MLLMs) but introduces critical backdoor risks via poisoned data. Existing defenses either rely on supervised signals or fail to generalize across diverse trigger types and modalities. In this work, we uncover a universal backdoor fingerprint-attention allocation divergence-where poisoned samples disrupt the balanced attention distribution across three functional components: system instructions, vision inputs, and user textual queries, regardless of trigger morphology. Motivated by this insight, "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2601.21692","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/2601.21692/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":"2601.21692","created_at":"2026-05-25T02:01:13.606223+00:00"},{"alias_kind":"arxiv_version","alias_value":"2601.21692v2","created_at":"2026-05-25T02:01:13.606223+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2601.21692","created_at":"2026-05-25T02:01:13.606223+00:00"},{"alias_kind":"pith_short_12","alias_value":"7PZ65OX2Y3D4","created_at":"2026-05-25T02:01:13.606223+00:00"},{"alias_kind":"pith_short_16","alias_value":"7PZ65OX2Y3D4XZ7I","created_at":"2026-05-25T02:01:13.606223+00:00"},{"alias_kind":"pith_short_8","alias_value":"7PZ65OX2","created_at":"2026-05-25T02:01:13.606223+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2605.15711","citing_title":"EntropyScan: Towards Model-level Backdoor Detection in LVLMs via Visual Attention Entropy","ref_index":30,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/7PZ65OX2Y3D4XZ7IKHJK5IVOA5","json":"https://pith.science/pith/7PZ65OX2Y3D4XZ7IKHJK5IVOA5.json","graph_json":"https://pith.science/api/pith-number/7PZ65OX2Y3D4XZ7IKHJK5IVOA5/graph.json","events_json":"https://pith.science/api/pith-number/7PZ65OX2Y3D4XZ7IKHJK5IVOA5/events.json","paper":"https://pith.science/paper/7PZ65OX2"},"agent_actions":{"view_html":"https://pith.science/pith/7PZ65OX2Y3D4XZ7IKHJK5IVOA5","download_json":"https://pith.science/pith/7PZ65OX2Y3D4XZ7IKHJK5IVOA5.json","view_paper":"https://pith.science/paper/7PZ65OX2","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2601.21692&json=true","fetch_graph":"https://pith.science/api/pith-number/7PZ65OX2Y3D4XZ7IKHJK5IVOA5/graph.json","fetch_events":"https://pith.science/api/pith-number/7PZ65OX2Y3D4XZ7IKHJK5IVOA5/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/7PZ65OX2Y3D4XZ7IKHJK5IVOA5/action/timestamp_anchor","attest_storage":"https://pith.science/pith/7PZ65OX2Y3D4XZ7IKHJK5IVOA5/action/storage_attestation","attest_author":"https://pith.science/pith/7PZ65OX2Y3D4XZ7IKHJK5IVOA5/action/author_attestation","sign_citation":"https://pith.science/pith/7PZ65OX2Y3D4XZ7IKHJK5IVOA5/action/citation_signature","submit_replication":"https://pith.science/pith/7PZ65OX2Y3D4XZ7IKHJK5IVOA5/action/replication_record"}},"created_at":"2026-05-25T02:01:13.606223+00:00","updated_at":"2026-05-25T02:01:13.606223+00:00"}