{"id":"W4385679821","doi":"10.1109/sp46215.2023.10179300","title":"Analyzing Leakage of Personally Identifiable Information in Language Models","year":2023,"lang":"en","type":"article","venue":"","topic":"Privacy-Preserving Technologies in Data","field":"Computer Science","cited_by":125,"is_retracted":false,"has_abstract":true,"ca_institutions":"University of Waterloo","funders":"","keywords":"Computer science; Information leakage; Leakage (economics); Natural language processing; Artificial intelligence; Internet privacy; Computer security","routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false,"invisible_to_affiliation_only":false},"retraction":null,"screen":null,"direct_labels":[],"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["open_science"],"consensus_categories":["open_science"],"category_scores_codex":[0.0005267673,0.00006069801,0.0001004628,0.0004816704,0.00002503608,0.00008492663,0.009026772,0.00005069618,0.00001306526],"category_scores_gemma":[0.003055353,0.0000580739,0.00002539127,0.001388857,0.00002365329,0.003175316,0.02294163,0.00009690972,0.00008984227],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00003216965,"about_ca_system_score_gemma":0.00002868507,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0003204558,"about_ca_topic_score_gemma":0.00005261603,"domain_scores_codex":[0.9992189,0.00001734337,0.0002274972,0.0001415615,0.0002023112,0.0001924123],"domain_scores_gemma":[0.9973573,0.00008357775,0.00006808162,0.002440855,0.00003378348,0.00001637178],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.00001503828,0.0001171355,0.008547585,0.0004718935,0.00007798475,0.0001057035,0.01123644,0.02868911,0.01707243,0.182013,0.5971792,0.1544744],"study_design_scores_gemma":[0.0001004717,0.000006881158,0.0007262522,0.00002088177,7.470318e-7,0.00000100871,0.0004404366,0.9241344,0.003939735,0.07049947,0.00006331741,0.00006636298],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.2332105,0.00004274392,0.7536468,0.004560935,0.0001137611,0.0001151884,0.00001105421,0.001097865,0.007201194],"genre_scores_gemma":[0.8781814,0.0000258062,0.1215034,0.00006254865,0.000003577346,0.000007394439,0.00002118398,0.000003238501,0.0001914432],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.8954453,"threshold_uncertainty_score":0.9963349,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.02843659952724117,"score_gpt":0.2731366741085866,"score_spread":0.2447000745813454,"validation_status":"score_only:v0-immature-baseline","note":"Baseline scores from an immature model (maturity gate not passed). Scores rank; they never assert a category."}}