{"id":"https://openalex.org/W7162435250","doi":"https://doi.org/10.48550/arxiv.2605.24879","title":"Efficient DP-SGD for LLMs with Randomized Clipping","display_name":"Efficient DP-SGD for LLMs with Randomized Clipping","publication_year":2026,"publication_date":"2026-05-24","ids":{"openalex":"https://openalex.org/W7162435250","doi":"https://doi.org/10.48550/arxiv.2605.24879"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2605.24879","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.24879","pdf_url":null,"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,"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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"},"type":"preprint","indexed_in":["datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://doi.org/10.48550/arxiv.2605.24879","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5024577194","display_name":"Enayat Ullah","orcid":"https://orcid.org/0000-0002-1395-4255"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ullah, Enayat","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5005856256","display_name":"Sai Aparna Aketi","orcid":"https://orcid.org/0000-0003-3446-0243"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Aketi, Sai Aparna","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5049261335","display_name":"Devansh Gupta","orcid":"https://orcid.org/0000-0001-9180-9611"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Gupta, Devansh","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5137056035","display_name":"Huanyu Zhang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhang, Huanyu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5041237539","display_name":"Meisam Razaviyayn","orcid":"https://orcid.org/0000-0003-4342-6661"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Razaviyayn, Meisam","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":0,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":false,"cited_by_count":0,"citation_normalized_percentile":null,"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":null,"last_page":null},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10764","display_name":"Privacy-Preserving Technologies in Data","score":0.6022999882698059,"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/T10764","display_name":"Privacy-Preserving Technologies in Data","score":0.6022999882698059,"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/T10028","display_name":"Topic Modeling","score":0.0843999981880188,"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/T11307","display_name":"Domain Adaptation and Few-Shot Learning","score":0.033799998462200165,"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/automatic-summarization","display_name":"Automatic summarization","score":0.6614000201225281},{"id":"https://openalex.org/keywords/context","display_name":"Context (archaeology)","score":0.5530999898910522},{"id":"https://openalex.org/keywords/differential-privacy","display_name":"Differential privacy","score":0.5412999987602234},{"id":"https://openalex.org/keywords/clipping","display_name":"Clipping (morphology)","score":0.5134000182151794},{"id":"https://openalex.org/keywords/memory-footprint","display_name":"Memory footprint","score":0.4675999879837036},{"id":"https://openalex.org/keywords/noise","display_name":"Noise (video)","score":0.3837999999523163},{"id":"https://openalex.org/keywords/sequence","display_name":"Sequence (biology)","score":0.34540000557899475},{"id":"https://openalex.org/keywords/overhead","display_name":"Overhead (engineering)","score":0.33090001344680786}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7457000017166138},{"id":"https://openalex.org/C170858558","wikidata":"https://www.wikidata.org/wiki/Q1394144","display_name":"Automatic summarization","level":2,"score":0.6614000201225281},{"id":"https://openalex.org/C2779343474","wikidata":"https://www.wikidata.org/wiki/Q3109175","display_name":"Context (archaeology)","level":2,"score":0.5530999898910522},{"id":"https://openalex.org/C23130292","wikidata":"https://www.wikidata.org/wiki/Q5275358","display_name":"Differential privacy","level":2,"score":0.5412999987602234},{"id":"https://openalex.org/C2776848632","wikidata":"https://www.wikidata.org/wiki/Q853463","display_name":"Clipping (morphology)","level":2,"score":0.5134000182151794},{"id":"https://openalex.org/C74912251","wikidata":"https://www.wikidata.org/wiki/Q6815727","display_name":"Memory footprint","level":2,"score":0.4675999879837036},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.4239000082015991},{"id":"https://openalex.org/C99498987","wikidata":"https://www.wikidata.org/wiki/Q2210247","display_name":"Noise (video)","level":3,"score":0.3837999999523163},{"id":"https://openalex.org/C2778112365","wikidata":"https://www.wikidata.org/wiki/Q3511065","display_name":"Sequence (biology)","level":2,"score":0.34540000557899475},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.3328999876976013},{"id":"https://openalex.org/C2779960059","wikidata":"https://www.wikidata.org/wiki/Q7113681","display_name":"Overhead (engineering)","level":2,"score":0.33090001344680786},{"id":"https://openalex.org/C12725497","wikidata":"https://www.wikidata.org/wiki/Q810247","display_name":"Baseline (sea)","level":2,"score":0.3167000114917755},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.28189998865127563},{"id":"https://openalex.org/C107673813","wikidata":"https://www.wikidata.org/wiki/Q812534","display_name":"Bayesian probability","level":2,"score":0.2818000018596649},{"id":"https://openalex.org/C191795146","wikidata":"https://www.wikidata.org/wiki/Q3878446","display_name":"Norm (philosophy)","level":2,"score":0.27709999680519104},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.27559998631477356},{"id":"https://openalex.org/C123201435","wikidata":"https://www.wikidata.org/wiki/Q456632","display_name":"Information privacy","level":2,"score":0.2734000086784363},{"id":"https://openalex.org/C26713055","wikidata":"https://www.wikidata.org/wiki/Q245962","display_name":"Implementation","level":2,"score":0.2712000012397766},{"id":"https://openalex.org/C206688291","wikidata":"https://www.wikidata.org/wiki/Q7617819","display_name":"Stochastic gradient descent","level":3,"score":0.26739999651908875},{"id":"https://openalex.org/C124101348","wikidata":"https://www.wikidata.org/wiki/Q172491","display_name":"Data mining","level":1,"score":0.2660999894142151},{"id":"https://openalex.org/C2780657452","wikidata":"https://www.wikidata.org/wiki/Q1193170","display_name":"Headset","level":2,"score":0.2605000138282776},{"id":"https://openalex.org/C28490314","wikidata":"https://www.wikidata.org/wiki/Q189436","display_name":"Speech recognition","level":1,"score":0.2603999972343445},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.259799987077713},{"id":"https://openalex.org/C75291252","wikidata":"https://www.wikidata.org/wiki/Q1315756","display_name":"TRACE (psycholinguistics)","level":2,"score":0.2583000063896179},{"id":"https://openalex.org/C28855332","wikidata":"https://www.wikidata.org/wiki/Q198099","display_name":"Quantization (signal processing)","level":2,"score":0.25600001215934753},{"id":"https://openalex.org/C63479239","wikidata":"https://www.wikidata.org/wiki/Q7353546","display_name":"Robustness (evolution)","level":3,"score":0.2515999972820282},{"id":"https://openalex.org/C38858127","wikidata":"https://www.wikidata.org/wiki/Q5441228","display_name":"Feed forward","level":2,"score":0.25049999356269836}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2605.24879","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.24879","pdf_url":null,"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,"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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"doi:10.48550/arxiv.2605.24879","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.24879","pdf_url":null,"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,"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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"},"sustainable_development_goals":[{"display_name":"Peace, Justice and strong institutions","id":"https://metadata.un.org/sdg/16","score":0.5941643118858337}],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Large":[0],"language":[1],"models":[2],"(LLMs)":[3],"are":[4],"trained":[5],"on":[6,42,145],"vast":[7],"datasets":[8],"that":[9,92,133,156],"may":[10],"contain":[11],"sensitive":[12],"information.":[13],"Differential":[14],"privacy":[15,23,34,130],"(DP),":[16],"the":[17,56,61,67,118],"de":[18],"facto":[19],"standard":[20],"for":[21,29],"formal":[22],"guarantees,":[24],"provides":[25],"a":[26,84,128],"principled":[27],"framework":[28],"training":[30,39],"LLMs":[31],"with":[32,47,89,139],"provable":[33],"protection.":[35],"However,":[36],"state-of-the-art":[37],"DP":[38],"implementations":[40],"rely":[41],"fast":[43],"gradient":[44,123],"clipping":[45,91],"techniques":[46],"memory":[48,94,119,164],"overhead":[49],"$O(B":[50],"\\min\\{T^2,":[51],"d^2\\})$,":[52],"where":[53],"$B$":[54],"is":[55,60,66],"batch":[57],"size,":[58],"$T$":[59],"sequence":[62],"length,":[63],"and":[64,77,95,108,152,165],"$d$":[65],"model":[68,75],"width.":[69],"This":[70],"becomes":[71],"prohibitive":[72],"as":[73],"both":[74],"size":[76],"context":[78],"length":[79],"grow.":[80],"We":[81,126],"propose":[82],"DP-SGD-RC,":[83],"novel":[85],"variant":[86],"of":[87,121],"DP-SGD":[88],"randomized":[90],"reduces":[93],"compute":[96,166],"complexity.":[97],"DP-SGD-RC":[98,134,157],"leverages":[99],"stochastic":[100],"trace":[101],"estimation":[102],"methods,":[103],"specifically":[104],"Hutchinson's":[105],"estimator[Hutchinson,":[106],"1989]":[107],"its":[109],"improved":[110],"variant,":[111],"Hutch++[Meyer":[112],"et":[113],"al.,":[114],"2021],":[115],"to":[116],"reduce":[117],"footprint":[120],"per-sample":[122],"norm":[124],"estimation.":[125],"provide":[127],"tight":[129],"analysis":[131],"showing":[132],"achieves":[135],"noise":[136],"multipliers":[137],"competitive":[138],"deterministic":[140],"clipping.":[141],"Experiments":[142],"fine-tuning":[143],"Llama~3.2-1B":[144],"long-context":[146],"benchmarks":[147],"spanning":[148],"classification,":[149],"question":[150],"answering,":[151],"summarization":[153],"tasks":[154],"demonstrate":[155],"matches":[158],"baseline":[159],"utility":[160],"while":[161],"significantly":[162],"reducing":[163],"requirements.":[167]},"counts_by_year":[],"updated_date":"2026-07-28T07:46:37.118299","created_date":"2026-05-27T00:00:00"}