{"id":"W2994896922","doi":"","title":"Thieves of Sesame Street: Model Extraction on BERT-based APIs","year":2020,"lang":"en","type":"article","venue":"arXiv (Cornell University)","topic":"Adversarial Robustness in Machine Learning","field":"Computer Science","cited_by":32,"is_retracted":false,"has_abstract":true,"ca_institutions":"University of Toronto","funders":"","keywords":"Computer science; Heuristics; Adversary; Language model; Set (abstract data type); Artificial intelligence; Task (project management); Exploit; Inference; Natural language processing; Computer security; Programming language","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":[],"consensus_categories":[],"category_scores_codex":[0.0001037231,0.000143679,0.0001666069,0.0001067028,0.0001031249,0.00002185474,0.0007476292,0.00008008069,0.00001968852],"category_scores_gemma":[0.00008963177,0.0001619262,0.00009789634,0.0006070522,0.0000618429,0.0005009075,0.0001631549,0.0002645408,0.00003145476],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0000581503,"about_ca_system_score_gemma":0.00008760661,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0000413607,"about_ca_topic_score_gemma":0.00000799384,"domain_scores_codex":[0.9989918,0.00009198407,0.0001276602,0.000489425,0.0001134175,0.000185683],"domain_scores_gemma":[0.9990937,0.0001529733,0.0001623894,0.0004024435,0.00006851602,0.0001200195],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.0000524401,0.000044257,0.001920779,0.00001484861,0.00001222637,0.00003652698,0.0001501677,0.8857291,0.0004300514,0.1110439,0.00006616255,0.0004995412],"study_design_scores_gemma":[0.0004912053,0.0001260945,0.0003193695,0.00001789622,0.00001596366,3.2767e-7,0.0000502536,0.9943042,0.001730827,0.002720369,0.0000655059,0.0001580303],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.1451611,0.00000401025,0.8525916,0.0004877858,0.00007318391,0.00007574084,0.000002591513,0.0001522505,0.001451745],"genre_scores_gemma":[0.9893546,0.000004279797,0.01006589,0.0003912905,0.00003351329,1.580187e-7,0.000002283078,0.00001072757,0.0001372774],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.8441935,"threshold_uncertainty_score":0.660316,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.08559918538337095,"score_gpt":0.2027448491080314,"score_spread":0.1171456637246604,"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."}}