{"id":"https://openalex.org/W3028525609","doi":"https://doi.org/10.1109/focs52979.2021.00098","title":"Feature Purification: How Adversarial Training Performs Robust Deep Learning","display_name":"Feature Purification: How Adversarial Training Performs Robust Deep Learning","alternate_titles":[],"publication_year":2022,"publication_date":"2022-02-01","ids":{"openalex":"https://openalex.org/W3028525609","doi":"https://doi.org/10.1109/focs52979.2021.00098","mag":"3028525609"},"language":"en","primary_location":{"id":"doi:10.1109/focs52979.2021.00098","doi":"https://doi.org/10.1109/focs52979.2021.00098","is_oa":false,"landing_page_url":"https://doi.org/10.1109/focs52979.2021.00098","pdf_url":null,"source":{"id":"https://openalex.org/S4210175768","display_name":"Annual Symposium on Foundations of Computer Science","issn_l":"0272-5428","issn":["0272-5428"],"is_oa":false,"is_in_doaj":false,"is_core":true,"listed_in":["cwts-core"],"host_organization":"https://openalex.org/P4310319808","host_organization_name":"Institute of Electrical and Electronics Engineers","host_organization_lineage":["https://openalex.org/P4310319808"],"host_organization_lineage_names":["Institute of Electrical and Electronics Engineers"],"type":"conference"},"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"2021 IEEE 62nd Annual Symposium on Foundations of Computer Science (FOCS)","raw_type":"proceedings-article"},"parent_works":[],"lineage":["https://openalex.org/W3028525609"],"child_works_count":0,"type":"conference-paper","is_traditional":true,"indexed_in":["arxiv","crossref","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2005.10190","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5048981295","display_name":"Zeyuan Allen-Zhu","orcid":"https://orcid.org/0000-0003-3002-089X","observed_orcids":["https://orcid.org/0000-0003-3002-089X"]},"institutions":[{"id":"https://openalex.org/I1290206253","display_name":"Microsoft (United States)","ror":"https://ror.org/00d0nc645","country_code":"US","type":"company","lineage":["https://openalex.org/I1290206253"]}],"countries":["US"],"is_corresponding":false,"contributor_roles":[],"raw_author_name":"Zeyuan Allen-Zhu","raw_affiliation_strings":["Microsoft Research Redmond, Redmond, WA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Microsoft Research Redmond, Redmond, WA","institution_ids":["https://openalex.org/I1290206253"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5101879892","display_name":"Yuanzhi Li","orcid":"https://orcid.org/0000-0002-4604-0468","observed_orcids":["https://orcid.org/0000-0002-4604-0468"]},"institutions":[{"id":"https://openalex.org/I74973139","display_name":"Carnegie Mellon University","ror":"https://ror.org/05x2bcf33","country_code":"US","type":"education","lineage":["https://openalex.org/I74973139"]}],"countries":["US"],"is_corresponding":false,"contributor_roles":[],"raw_author_name":"Yuanzhi Li","raw_affiliation_strings":["Carnegie Mellon University, Pittsburgh, PA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Carnegie Mellon University, Pittsburgh, PA","institution_ids":["https://openalex.org/I74973139"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":2,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":10.5855,"has_fulltext":true,"cited_by_count":211,"citation_normalized_percentile":{"value":0.98944381,"is_in_top_1_percent":false,"is_in_top_10_percent":true},"cited_by_percentile_year":{"min":87,"max":100},"biblio":{"volume":null,"issue":null,"first_page":"977","last_page":"988"},"is_retracted":false,"is_pulled":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T11689","display_name":"Adversarial Robustness in Machine Learning","score":1.0,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},"topics":[{"id":"https://openalex.org/T11689","display_name":"Adversarial Robustness in Machine Learning","score":1.0,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T11612","display_name":"Stochastic Gradient Optimization Techniques","score":0.0,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T10764","display_name":"Privacy-Preserving Technologies in Data","score":0.0,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}}],"keywords":[{"id":"https://openalex.org/keywords/adversarial-training","display_name":"adversarial training","score":0.8980000019073486},{"id":"https://openalex.org/keywords/adversarial-examples","display_name":"adversarial examples","score":0.7639999985694885},{"id":"https://openalex.org/keywords/adversarial-robustness","display_name":"adversarial robustness","score":0.8349999785423279},{"id":"https://openalex.org/keywords/two-layer-neural-networks","display_name":"two-layer neural networks","score":0.75},{"id":"https://openalex.org/keywords/relu-activation-function","display_name":"ReLU activation","score":0.9769999980926514},{"id":"https://openalex.org/keywords/gradient-descent","display_name":"gradient descent","score":0.5809999704360962},{"id":"https://openalex.org/keywords/fgm","display_name":"FGM","score":0.44699999690055847},{"id":"https://openalex.org/keywords/complexity-lower-bounds","display_name":"complexity lower bounds","score":0.7080000042915344},{"id":"https://openalex.org/keywords/neural-tangent-kernel","display_name":"neural tangent kernel","score":0.9570000171661377},{"id":"https://openalex.org/keywords/linear-classifiers","display_name":"linear classifiers","score":0.5059999823570251},{"id":"https://openalex.org/keywords/low-degree-polynomials","display_name":"low-degree polynomials","score":0.7990000247955322}],"concepts":[{"id":"https://openalex.org/C37736160","wikidata":"https://www.wikidata.org/wiki/Q1801315","display_name":"Adversarial system","level":2,"score":0.8540879487991333},{"id":"https://openalex.org/C50644808","wikidata":"https://www.wikidata.org/wiki/Q192776","display_name":"Artificial neural network","level":2,"score":0.6911234855651855},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.6308290362358093},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6206983923912048},{"id":"https://openalex.org/C2777211547","wikidata":"https://www.wikidata.org/wiki/Q17141490","display_name":"Training (meteorology)","level":2,"score":0.48412618041038513},{"id":"https://openalex.org/C2776401178","wikidata":"https://www.wikidata.org/wiki/Q12050496","display_name":"Feature (linguistics)","level":2,"score":0.45879504084587097},{"id":"https://openalex.org/C153258448","wikidata":"https://www.wikidata.org/wiki/Q1199743","display_name":"Gradient descent","level":3,"score":0.4527660608291626},{"id":"https://openalex.org/C149728462","wikidata":"https://www.wikidata.org/wiki/Q751319","display_name":"Minimax","level":2,"score":0.4445607662200928},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.4356442987918854},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.43155306577682495},{"id":"https://openalex.org/C2984842247","wikidata":"https://www.wikidata.org/wiki/Q197536","display_name":"Deep neural networks","level":3,"score":0.4307190179824829},{"id":"https://openalex.org/C108583219","wikidata":"https://www.wikidata.org/wiki/Q197536","display_name":"Deep learning","level":2,"score":0.42230430245399475},{"id":"https://openalex.org/C177918212","wikidata":"https://www.wikidata.org/wiki/Q803623","display_name":"Perturbation (astronomy)","level":2,"score":0.4218827784061432},{"id":"https://openalex.org/C153180895","wikidata":"https://www.wikidata.org/wiki/Q7148389","display_name":"Pattern recognition (psychology)","level":2,"score":0.3256082832813263},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.2999802231788635},{"id":"https://openalex.org/C126255220","wikidata":"https://www.wikidata.org/wiki/Q141495","display_name":"Mathematical optimization","level":1,"score":0.24515065550804138},{"id":"https://openalex.org/C153294291","wikidata":"https://www.wikidata.org/wiki/Q25261","display_name":"Meteorology","level":1,"score":0.0},{"id":"https://openalex.org/C138885662","wikidata":"https://www.wikidata.org/wiki/Q5891","display_name":"Philosophy","level":0,"score":0.0},{"id":"https://openalex.org/C62520636","wikidata":"https://www.wikidata.org/wiki/Q944","display_name":"Quantum mechanics","level":1,"score":0.0},{"id":"https://openalex.org/C121332964","wikidata":"https://www.wikidata.org/wiki/Q413","display_name":"Physics","level":0,"score":0.0},{"id":"https://openalex.org/C41895202","wikidata":"https://www.wikidata.org/wiki/Q8162","display_name":"Linguistics","level":1,"score":0.0}],"mesh":[],"locations_count":3,"locations":[{"id":"doi:10.1109/focs52979.2021.00098","doi":"https://doi.org/10.1109/focs52979.2021.00098","is_oa":false,"landing_page_url":"https://doi.org/10.1109/focs52979.2021.00098","pdf_url":null,"source":{"id":"https://openalex.org/S4210175768","display_name":"Annual Symposium on Foundations of Computer Science","issn_l":"0272-5428","issn":["0272-5428"],"is_oa":false,"is_in_doaj":false,"is_core":true,"listed_in":["cwts-core"],"host_organization":"https://openalex.org/P4310319808","host_organization_name":"Institute of Electrical and Electronics Engineers","host_organization_lineage":["https://openalex.org/P4310319808"],"host_organization_lineage_names":["Institute of Electrical and Electronics Engineers"],"type":"conference"},"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"2021 IEEE 62nd Annual Symposium on Foundations of Computer Science (FOCS)","raw_type":"proceedings-article"},{"id":"pmh:oai:arXiv.org:2005.10190","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2005.10190","pdf_url":"https://arxiv.org/pdf/2005.10190","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2005.10190","doi":"https://doi.org/10.48550/arxiv.2005.10190","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2005.10190","pdf_url":"https://arxiv.org/pdf/2005.10190","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2005.10190","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2005.10190","pdf_url":"https://arxiv.org/pdf/2005.10190","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W3028525609.pdf","grobid_xml":"https://content.openalex.org/works/W3028525609.grobid-xml"},"referenced_works_count":126,"referenced_works":["https://openalex.org/W9657784","https://openalex.org/W1673923490","https://openalex.org/W1836465849","https://openalex.org/W1915485278","https://openalex.org/W1932198206","https://openalex.org/W1945616565","https://openalex.org/W2005876975","https://openalex.org/W2014485295","https://openalex.org/W2034683677","https://openalex.org/W2053885701","https://openalex.org/W2058670155","https://openalex.org/W2060775054","https://openalex.org/W2074376560","https://openalex.org/W2085400714","https://openalex.org/W2097018403","https://openalex.org/W2105459687","https://openalex.org/W2105464873","https://openalex.org/W2106334384","https://openalex.org/W2109999391","https://openalex.org/W2113606819","https://openalex.org/W2116216716","https://openalex.org/W2117508909","https://openalex.org/W2117822564","https://openalex.org/W2118103795","https://openalex.org/W2119385818","https://openalex.org/W2143612262","https://openalex.org/W2147154704","https://openalex.org/W2163605009","https://openalex.org/W2194775991","https://openalex.org/W2257979135","https://openalex.org/W2295508206","https://openalex.org/W2399994860","https://openalex.org/W2404068508","https://openalex.org/W2407000478","https://openalex.org/W2418098761","https://openalex.org/W2513314332","https://openalex.org/W2518186251","https://openalex.org/W2547369620","https://openalex.org/W2557139899","https://openalex.org/W2557449848","https://openalex.org/W2573849764","https://openalex.org/W2591714514","https://openalex.org/W2593709294","https://openalex.org/W2604084771","https://openalex.org/W2610386774","https://openalex.org/W2614119628","https://openalex.org/W2625063094","https://openalex.org/W2626994141","https://openalex.org/W2640329709","https://openalex.org/W2739160792","https://openalex.org/W2765233338","https://openalex.org/W2765384636","https://openalex.org/W2765701314","https://openalex.org/W2765861987","https://openalex.org/W2766371994","https://openalex.org/W2783692467","https://openalex.org/W2787480667","https://openalex.org/W2787496614","https://openalex.org/W2787708942","https://openalex.org/W2789048837","https://openalex.org/W2799032899","https://openalex.org/W2800987155","https://openalex.org/W2803732607","https://openalex.org/W2805003733","https://openalex.org/W2808958252","https://openalex.org/W2809090039","https://openalex.org/W2886067286","https://openalex.org/W2890689004","https://openalex.org/W2891572479","https://openalex.org/W2894604724","https://openalex.org/W2897355816","https://openalex.org/W2898324627","https://openalex.org/W2898871740","https://openalex.org/W2899748887","https://openalex.org/W2899790086","https://openalex.org/W2900103278","https://openalex.org/W2900959181","https://openalex.org/W2902031012","https://openalex.org/W2911867426","https://openalex.org/W2913266441","https://openalex.org/W2913473169","https://openalex.org/W2941057241","https://openalex.org/W2942052807","https://openalex.org/W2942630857","https://openalex.org/W2945404573","https://openalex.org/W2945793108","https://openalex.org/W2946840143","https://openalex.org/W2949117887","https://openalex.org/W2949769151","https://openalex.org/W2950787360","https://openalex.org/W2951576642","https://openalex.org/W2951735139","https://openalex.org/W2951884431","https://openalex.org/W2952817981","https://openalex.org/W2962698540","https://openalex.org/W2962710014","https://openalex.org/W2962939986","https://openalex.org/W2962959718","https://openalex.org/W2963344237","https://openalex.org/W2963353335","https://openalex.org/W2963519230","https://openalex.org/W2963774180","https://openalex.org/W2964137095","https://openalex.org/W2964153729","https://openalex.org/W2966363981","https://openalex.org/W2970241199","https://openalex.org/W2970388773","https://openalex.org/W2970443625","https://openalex.org/W2972392510","https://openalex.org/W2996852294","https://openalex.org/W3000127803","https://openalex.org/W3007366066","https://openalex.org/W3010476364","https://openalex.org/W3034175346","https://openalex.org/W3042003204","https://openalex.org/W3098560605","https://openalex.org/W3103836116","https://openalex.org/W3130725574","https://openalex.org/W3168867926","https://openalex.org/W4229494842","https://openalex.org/W4240805545","https://openalex.org/W4289147935","https://openalex.org/W4383859664","https://openalex.org/W6729433981","https://openalex.org/W6740096615","https://openalex.org/W7099032813"],"related_works":["https://openalex.org/W4403444251","https://openalex.org/W4385880757","https://openalex.org/W4411336749","https://openalex.org/W2785466457","https://openalex.org/W3010028534","https://openalex.org/W4391377077","https://openalex.org/W2979524290","https://openalex.org/W7131826980","https://openalex.org/W4391418199","https://openalex.org/W4312646667"],"abstract_inverted_index":{"Despite":[0],"the":[1,24,28,39,62,65,71,79,83,95,114,153,158,163,180,215,239],"empirical":[2,199],"success":[3],"of":[4,30,61,64,67,73,86,94,97,155,190,214,250,265],"using":[5,140],"adversarial":[6,14,31,35,68,98,188],"training":[7,36,84,99,132,170],"to":[8,38,42,101,105,117,152,186,260],"defend":[9,247],"deep":[10],"learning":[11],"models":[12,230],"against":[13,211,248],"perturbations,":[15,32],"so":[16],"far,":[17],"it":[18],"still":[19],"remains":[20],"rather":[21],"unclear":[22],"what":[23,34,256],"principles":[25],"are":[26,258],"behind":[27],"existence":[29,66],"and":[33,90,121],"does":[37],"neural":[40,88,135,172,240],"network":[41,136,173],"remove":[43,102],"them.":[44],"In":[45],"this":[46,119,147,244,251,266],"paper,":[47],"we":[48,53,58,150,219],"present":[49,110],"a":[50,87,122,133,171,222],"principle":[51],"that":[52,126,162,227],"call":[54],"feature":[55],"purification,":[56],"where":[57],"show":[59],"one":[60,93],"causes":[63],"examples":[69],"is":[70,100,183],"accumulation":[72],"certain":[74,128],"small":[75,187],"dense":[76],"mixtures":[77,104],"in":[78,206],"hidden":[80,107],"weights":[81],"during":[82],"process":[85],"network;":[89],"more":[91],"importantly,":[92],"goals":[96],"such":[103,202,231],"purify":[106],"weights.":[108],"We":[109],"both":[111],"experiments":[112],"on":[113],"CIFAR-10":[115],"dataset":[116],"illustrate":[118],"principle,":[120],"theoretical":[123],"result":[124],"proving":[125,161],"for":[127,169,243],"natural":[129],"classification":[130],"tasks,":[131],"two-layer":[134],"with":[137,174,197],"ReLU":[138,175],"activation":[139],"randomly":[141],"initialized":[142],"gradient":[143],"descent":[144],"indeed":[145,184],"satisfies":[146],"principle.":[148],"Technically,":[149],"give,":[151],"best":[154],"our":[156],"knowledge,":[157],"first":[159],"re-sult":[160],"following":[164],"two":[165],"can":[166,205,268],"hold":[167],"simultaneously":[168],"activation.":[176],"(1)":[177],"Training":[178],"over":[179],"original":[181],"data":[182],"non-robust":[185],"perturbations":[189,213,249],"some":[191],"radius.":[192,217],"(2)":[193],"Adversarial":[194],"training,":[195],"even":[196,238],"an":[198],"perturbation":[200],"algorithm":[201],"as":[203,232],"FGM,":[204],"fact":[207],"be":[208,269],"provably":[209],"robust":[210],"any":[212],"same":[216,252],"Finally,":[218],"also":[220],"prove":[221],"complexity":[223,229],"lower":[224],"bound,":[225],"showing":[226],"low":[228],"linear":[233],"classifiers,":[234],"low-degree":[235],"polynomials,":[236],"or":[237],"tangent":[241],"kernel":[242],"network,":[245],"cannot":[246],"radius,":[253],"no":[254],"matter":[255],"algorithms":[257],"used":[259],"train":[261],"them.11The":[262],"full":[263],"version":[264],"paper":[267],"found":[270],"at":[271],"https://arxiv.org/abs/-2005.l0190.":[272]},"cited_by_api_url":"https://api.openalex.org/works?filter=cites:W3028525609","counts_by_year":[{"year":2026,"cited_by_count":19},{"year":2025,"cited_by_count":45},{"year":2024,"cited_by_count":39},{"year":2023,"cited_by_count":41},{"year":2022,"cited_by_count":31},{"year":2021,"cited_by_count":20},{"year":2020,"cited_by_count":14},{"year":2019,"cited_by_count":1},{"year":2018,"cited_by_count":1}],"updated_date":"2026-11-20T13:11:24.684918","created_date":"2020-05-29T00:00:00"}
