{"id":"https://openalex.org/W7162549559","doi":"https://doi.org/10.48550/arxiv.2605.26574","title":"GradSentry: Gradient Spectral Entropy for Backdoor Sample Filtering in Large Language Model Fine-Tuning","display_name":"GradSentry: Gradient Spectral Entropy for Backdoor Sample Filtering in Large Language Model Fine-Tuning","publication_year":2026,"publication_date":"2026-05-26","ids":{"openalex":"https://openalex.org/W7162549559","doi":"https://doi.org/10.48550/arxiv.2605.26574"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2605.26574","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.26574","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"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":null,"license_id":null,"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.2605.26574","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5137128793","display_name":"Haodong Zhao","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhao, Haodong","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5137173997","display_name":"Tianyi Xu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Xu, Tianyi","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5137154919","display_name":"Tianhang Zhao","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhao, Tianhang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5137126796","display_name":"Zhuosheng Zhang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhang, Zhuosheng","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5137141017","display_name":"Gongshen Liu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Liu, Gongshen","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/T11689","display_name":"Adversarial Robustness in Machine Learning","score":0.8877000212669373,"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":0.8877000212669373,"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/T12262","display_name":"Hate Speech and Cyberbullying Detection","score":0.023000000044703484,"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":0.015300000086426735,"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/backdoor","display_name":"Backdoor","score":0.9868000149726868},{"id":"https://openalex.org/keywords/pairwise-comparison","display_name":"Pairwise comparison","score":0.5723000168800354},{"id":"https://openalex.org/keywords/entropy","display_name":"Entropy (arrow of time)","score":0.5403000116348267},{"id":"https://openalex.org/keywords/cluster-analysis","display_name":"Cluster analysis","score":0.4618000090122223},{"id":"https://openalex.org/keywords/sample","display_name":"Sample (material)","score":0.4392000138759613},{"id":"https://openalex.org/keywords/key","display_name":"Key (lock)","score":0.40549999475479126}],"concepts":[{"id":"https://openalex.org/C2781045450","wikidata":"https://www.wikidata.org/wiki/Q254569","display_name":"Backdoor","level":2,"score":0.9868000149726868},{"id":"https://openalex.org/C184898388","wikidata":"https://www.wikidata.org/wiki/Q1435712","display_name":"Pairwise comparison","level":2,"score":0.5723000168800354},{"id":"https://openalex.org/C106301342","wikidata":"https://www.wikidata.org/wiki/Q4117933","display_name":"Entropy (arrow of time)","level":2,"score":0.5403000116348267},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.49880000948905945},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.46880000829696655},{"id":"https://openalex.org/C73555534","wikidata":"https://www.wikidata.org/wiki/Q622825","display_name":"Cluster analysis","level":2,"score":0.4618000090122223},{"id":"https://openalex.org/C198531522","wikidata":"https://www.wikidata.org/wiki/Q485146","display_name":"Sample (material)","level":2,"score":0.4392000138759613},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.40549999475479126},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.36160001158714294},{"id":"https://openalex.org/C9679016","wikidata":"https://www.wikidata.org/wiki/Q1417473","display_name":"Principle of maximum entropy","level":2,"score":0.3463999927043915},{"id":"https://openalex.org/C124101348","wikidata":"https://www.wikidata.org/wiki/Q172491","display_name":"Data mining","level":1,"score":0.3384999930858612},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.3294999897480011},{"id":"https://openalex.org/C153180895","wikidata":"https://www.wikidata.org/wiki/Q7148389","display_name":"Pattern recognition (psychology)","level":2,"score":0.3154999911785126},{"id":"https://openalex.org/C2779960059","wikidata":"https://www.wikidata.org/wiki/Q7113681","display_name":"Overhead (engineering)","level":2,"score":0.29409998655319214},{"id":"https://openalex.org/C105611402","wikidata":"https://www.wikidata.org/wiki/Q2976589","display_name":"Spectral clustering","level":3,"score":0.2752000093460083},{"id":"https://openalex.org/C2776401178","wikidata":"https://www.wikidata.org/wiki/Q12050496","display_name":"Feature (linguistics)","level":2,"score":0.2614000141620636}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2605.26574","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.26574","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"doi:10.48550/arxiv.2605.26574","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.26574","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Fine-tuning":[0],"Large":[1],"Language":[2],"Models":[3],"with":[4,62],"untrusted":[5],"data":[6,27],"exposes":[7],"models":[8],"to":[9,67],"backdoor":[10,41,73,157],"attacks,":[11],"where":[12],"poisoned":[13,58],"samples":[14,59],"cause":[15],"targeted":[16],"misbehavior.":[17],"Existing":[18],"sample-filtering":[19],"defenses":[20],"rely":[21],"on":[22,46,142],"clustering,":[23,122],"which":[24,112],"requires":[25,120],"sufficient":[26],"and":[28,83,102,130,146],"can":[29],"fail":[30],"at":[31,162],"extreme":[32],"poison":[33,127],"ratios.":[34],"We":[35],"propose":[36],"GradSentry":[37,70,119],"({Grad}ient":[38],"{Sentry}),":[39],"a":[40],"sample":[42,81,137],"filtering":[43],"method":[44,90],"based":[45],"the":[47,106,151],"spectral":[48,64,154],"entropy":[49,65,155],"of":[50,111,153],"per-sample":[51,76],"gradients.":[52],"Our":[53],"key":[54],"finding":[55],"is":[56,91,160],"that":[57],"produce":[60],"gradients":[61],"higher":[63],"compared":[66],"clean":[68],"samples.":[69],"captures":[71],"output-altering":[72],"signatures":[74],"using":[75],"gradient":[77,107],"spectra,":[78],"avoiding":[79],"pairwise":[80],"comparisons":[82],"clustering":[84],"during":[85,117],"feature":[86],"construction.":[87],"Importantly,":[88],"our":[89],"training-agnostic:":[92],"it":[93],"works":[94],"for":[95,138,156],"both":[96],"parameter-efficient":[97],"fine-tuning":[98],"methods":[99],"like":[100],"LoRA":[101],"full-parameter":[103],"tuning,":[104],"as":[105],"analysis":[108],"operates":[109,123],"independently":[110],"parameters":[113],"are":[114],"being":[115],"updated":[116],"training.":[118],"no":[121],"effectively":[124],"across":[125],"all":[126],"ratios":[128],"(1%--90%),":[129],"introduces":[131],"minimal":[132],"computational":[133],"overhead":[134],"(20-50ms":[135],"per":[136],"7B":[139],"model).":[140],"Evaluation":[141],"four":[143,147],"QA":[144],"datasets":[145],"attack":[148],"types":[149],"demonstrates":[150],"effectiveness":[152],"detection.":[158],"Code":[159],"available":[161],"https://github.com/dongdongzhaoUP/GradSentry.":[163]},"counts_by_year":[],"updated_date":"2026-07-01T06:00:48.157686","created_date":"2026-05-28T00:00:00"}