{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:YJI57KEFZ6XFRDEC444URU57WB","short_pith_number":"pith:YJI57KEF","schema_version":"1.0","canonical_sha256":"c251dfa885cfae588c82e73948d3bfb04badac45e8cad83b7827f659ad4eb2ab","source":{"kind":"arxiv","id":"2506.03234","version":1},"attestation_state":"computed","paper":{"title":"BadReward: Clean-Label Poisoning of Reward Models in Text-to-Image RLHF","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CR"],"primary_cat":"cs.LG","authors_text":"Cong Wang, Hongwei Yao, Kaiwen Duan, Tong Qiao, Yufei Chen, Zhan Qin, Ziyun Li","submitted_at":"2025-06-03T16:01:04Z","abstract_excerpt":"Reinforcement Learning from Human Feedback (RLHF) is crucial for aligning text-to-image (T2I) models with human preferences. However, RLHF's feedback mechanism also opens new pathways for adversaries. This paper demonstrates the feasibility of hijacking T2I models by poisoning a small fraction of preference data with natural-appearing examples. Specifically, we propose BadReward, a stealthy clean-label poisoning attack targeting the reward model in multi-modal RLHF. BadReward operates by inducing feature collisions between visually contradicted preference data instances, thereby corrupting the"},"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":"2506.03234","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-06-03T16:01:04Z","cross_cats_sorted":["cs.AI","cs.CR"],"title_canon_sha256":"7e314dbb397e7bc65995591d8ff09844964eb6d253e2f9788eaeffeb9c588083","abstract_canon_sha256":"044f2875ca8ea1cb43e992b200591f661fee26dd93b341872685260b59a5dbae"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:15:48.215447Z","signature_b64":"CNR2NI6KGyI6Pfh8na65QkIIxdMrUEgXYxbU9pL44S3kpnpk4Ou0XTWmToCyYEmdGUhdGczjJkHtmdArMeq9Cw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"c251dfa885cfae588c82e73948d3bfb04badac45e8cad83b7827f659ad4eb2ab","last_reissued_at":"2026-07-05T11:15:48.214947Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:15:48.214947Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"BadReward: Clean-Label Poisoning of Reward Models in Text-to-Image RLHF","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.CR"],"primary_cat":"cs.LG","authors_text":"Cong Wang, Hongwei Yao, Kaiwen Duan, Tong Qiao, Yufei Chen, Zhan Qin, Ziyun Li","submitted_at":"2025-06-03T16:01:04Z","abstract_excerpt":"Reinforcement Learning from Human Feedback (RLHF) is crucial for aligning text-to-image (T2I) models with human preferences. However, RLHF's feedback mechanism also opens new pathways for adversaries. This paper demonstrates the feasibility of hijacking T2I models by poisoning a small fraction of preference data with natural-appearing examples. Specifically, we propose BadReward, a stealthy clean-label poisoning attack targeting the reward model in multi-modal RLHF. BadReward operates by inducing feature collisions between visually contradicted preference data instances, thereby corrupting the"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2506.03234","kind":"arxiv","version":1},"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/2506.03234/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":"2506.03234","created_at":"2026-07-05T11:15:48.215002+00:00"},{"alias_kind":"arxiv_version","alias_value":"2506.03234v1","created_at":"2026-07-05T11:15:48.215002+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2506.03234","created_at":"2026-07-05T11:15:48.215002+00:00"},{"alias_kind":"pith_short_12","alias_value":"YJI57KEFZ6XF","created_at":"2026-07-05T11:15:48.215002+00:00"},{"alias_kind":"pith_short_16","alias_value":"YJI57KEFZ6XFRDEC","created_at":"2026-07-05T11:15:48.215002+00:00"},{"alias_kind":"pith_short_8","alias_value":"YJI57KEF","created_at":"2026-07-05T11:15:48.215002+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":3,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.09711","citing_title":"Proxy Reward Internalization and Mechanistic Exploitation: A Learned Precursor to Reward Hacking and Its Generalization","ref_index":295,"is_internal_anchor":false},{"citing_arxiv_id":"2605.10937","citing_title":"Power Reinforcement Post-Training of Text-to-Image Models with Super-Linear Advantage Shaping","ref_index":8,"is_internal_anchor":false},{"citing_arxiv_id":"2604.13602","citing_title":"Reward Hacking in the Era of Large Models: Mechanisms, Emergent Misalignment, Challenges","ref_index":221,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/YJI57KEFZ6XFRDEC444URU57WB","json":"https://pith.science/pith/YJI57KEFZ6XFRDEC444URU57WB.json","graph_json":"https://pith.science/api/pith-number/YJI57KEFZ6XFRDEC444URU57WB/graph.json","events_json":"https://pith.science/api/pith-number/YJI57KEFZ6XFRDEC444URU57WB/events.json","paper":"https://pith.science/paper/YJI57KEF"},"agent_actions":{"view_html":"https://pith.science/pith/YJI57KEFZ6XFRDEC444URU57WB","download_json":"https://pith.science/pith/YJI57KEFZ6XFRDEC444URU57WB.json","view_paper":"https://pith.science/paper/YJI57KEF","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2506.03234&json=true","fetch_graph":"https://pith.science/api/pith-number/YJI57KEFZ6XFRDEC444URU57WB/graph.json","fetch_events":"https://pith.science/api/pith-number/YJI57KEFZ6XFRDEC444URU57WB/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/YJI57KEFZ6XFRDEC444URU57WB/action/timestamp_anchor","attest_storage":"https://pith.science/pith/YJI57KEFZ6XFRDEC444URU57WB/action/storage_attestation","attest_author":"https://pith.science/pith/YJI57KEFZ6XFRDEC444URU57WB/action/author_attestation","sign_citation":"https://pith.science/pith/YJI57KEFZ6XFRDEC444URU57WB/action/citation_signature","submit_replication":"https://pith.science/pith/YJI57KEFZ6XFRDEC444URU57WB/action/replication_record"}},"created_at":"2026-07-05T11:15:48.215002+00:00","updated_at":"2026-07-05T11:15:48.215002+00:00"}