{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2021:QQXQN67SCKJ36AMMDGKOLPVMAH","short_pith_number":"pith:QQXQN67S","schema_version":"1.0","canonical_sha256":"842f06fbf21293bf018c1994e5beac01ff59f96d176ee7834f3da0427492f754","source":{"kind":"arxiv","id":"2106.09667","version":2},"attestation_state":"computed","paper":{"title":"Poisoning and Backdooring Contrastive Learning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Andreas Terzis, Nicholas Carlini","submitted_at":"2021-06-17T17:20:45Z","abstract_excerpt":"Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, ar"},"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":"2106.09667","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2021-06-17T17:20:45Z","cross_cats_sorted":[],"title_canon_sha256":"c986a1824022a058a2f6c64ef8f2ee472a8731600285e502cdd9b09f6de963fa","abstract_canon_sha256":"7421ac9a1c9a80c55768ae7c2fc207940429ab1631f50b13eb749a919b4cc4f6"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T04:08:57.481719Z","signature_b64":"4LPNzGWLO/K1u49RDVsZMuj73N3Jb5x12FH7fMuKpTTzg2K3bDDja2zhLjoEG+OUNK0NQ18O12zaJukPbO3xBw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"842f06fbf21293bf018c1994e5beac01ff59f96d176ee7834f3da0427492f754","last_reissued_at":"2026-07-05T04:08:57.481273Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T04:08:57.481273Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Poisoning and Backdooring Contrastive Learning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Andreas Terzis, Nicholas Carlini","submitted_at":"2021-06-17T17:20:45Z","abstract_excerpt":"Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, ar"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2106.09667","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/2106.09667/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":"2106.09667","created_at":"2026-07-05T04:08:57.481337+00:00"},{"alias_kind":"arxiv_version","alias_value":"2106.09667v2","created_at":"2026-07-05T04:08:57.481337+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2106.09667","created_at":"2026-07-05T04:08:57.481337+00:00"},{"alias_kind":"pith_short_12","alias_value":"QQXQN67SCKJ3","created_at":"2026-07-05T04:08:57.481337+00:00"},{"alias_kind":"pith_short_16","alias_value":"QQXQN67SCKJ36AMM","created_at":"2026-07-05T04:08:57.481337+00:00"},{"alias_kind":"pith_short_8","alias_value":"QQXQN67S","created_at":"2026-07-05T04:08:57.481337+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":5,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2607.07907","citing_title":"Multimodal Unlearning Across Vision, Language, Video, and Audio: Survey of Methods, Datasets, and Benchmarks","ref_index":289,"is_internal_anchor":true},{"citing_arxiv_id":"2305.17493","citing_title":"The Curse of Recursion: Training on Generated Data Makes Models Forget","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2109.13916","citing_title":"Unsolved Problems in ML Safety","ref_index":29,"is_internal_anchor":false},{"citing_arxiv_id":"2604.27781","citing_title":"The Grand Software Supply Chain of AI Systems","ref_index":2,"is_internal_anchor":false},{"citing_arxiv_id":"2604.05809","citing_title":"Stealthy and Adjustable Text-Guided Backdoor Attacks on Multimodal Pretrained Models","ref_index":25,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/QQXQN67SCKJ36AMMDGKOLPVMAH","json":"https://pith.science/pith/QQXQN67SCKJ36AMMDGKOLPVMAH.json","graph_json":"https://pith.science/api/pith-number/QQXQN67SCKJ36AMMDGKOLPVMAH/graph.json","events_json":"https://pith.science/api/pith-number/QQXQN67SCKJ36AMMDGKOLPVMAH/events.json","paper":"https://pith.science/paper/QQXQN67S"},"agent_actions":{"view_html":"https://pith.science/pith/QQXQN67SCKJ36AMMDGKOLPVMAH","download_json":"https://pith.science/pith/QQXQN67SCKJ36AMMDGKOLPVMAH.json","view_paper":"https://pith.science/paper/QQXQN67S","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2106.09667&json=true","fetch_graph":"https://pith.science/api/pith-number/QQXQN67SCKJ36AMMDGKOLPVMAH/graph.json","fetch_events":"https://pith.science/api/pith-number/QQXQN67SCKJ36AMMDGKOLPVMAH/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/QQXQN67SCKJ36AMMDGKOLPVMAH/action/timestamp_anchor","attest_storage":"https://pith.science/pith/QQXQN67SCKJ36AMMDGKOLPVMAH/action/storage_attestation","attest_author":"https://pith.science/pith/QQXQN67SCKJ36AMMDGKOLPVMAH/action/author_attestation","sign_citation":"https://pith.science/pith/QQXQN67SCKJ36AMMDGKOLPVMAH/action/citation_signature","submit_replication":"https://pith.science/pith/QQXQN67SCKJ36AMMDGKOLPVMAH/action/replication_record"}},"created_at":"2026-07-05T04:08:57.481337+00:00","updated_at":"2026-07-05T04:08:57.481337+00:00"}