{"id":"https://openalex.org/W7163164928","doi":"https://doi.org/10.48550/arxiv.2606.00160","title":"DataShield: Safety-degrading Data Filtering for LLM Benign Instruction Fine-Tuning","display_name":"DataShield: Safety-degrading Data Filtering for LLM Benign Instruction Fine-Tuning","publication_year":2026,"publication_date":"2026-05-29","ids":{"openalex":"https://openalex.org/W7163164928","doi":"https://doi.org/10.48550/arxiv.2606.00160"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2606.00160","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.00160","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.2606.00160","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5100778483","display_name":"Junbo Zhang","orcid":"https://orcid.org/0000-0003-2737-0821"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhang, Junbo","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5137706775","display_name":"Qianli Zhou","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhou, Qianli","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5137622576","display_name":"Xinyang Deng","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Deng, Xinyang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5137699457","display_name":"Wen Jiang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Jiang, Wen","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5137628456","display_name":"Jie Pan","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Pan, Jie","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5137622980","display_name":"Jinbiao Zhu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhu, Jinbiao","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":0.9429000020027161,"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":0.9429000020027161,"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/T11241","display_name":"Advanced Malware Detection Techniques","score":0.01119999960064888,"subfield":{"id":"https://openalex.org/subfields/1711","display_name":"Signal Processing"},"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":0.007799999788403511,"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/intuition","display_name":"Intuition","score":0.5523999929428101},{"id":"https://openalex.org/keywords/key","display_name":"Key (lock)","score":0.4668000042438507},{"id":"https://openalex.org/keywords/training-set","display_name":"Training set","score":0.3959999978542328},{"id":"https://openalex.org/keywords/noise","display_name":"Noise (video)","score":0.3831000030040741},{"id":"https://openalex.org/keywords/labeled-data","display_name":"Labeled data","score":0.3806000053882599},{"id":"https://openalex.org/keywords/code","display_name":"Code (set theory)","score":0.357699990272522},{"id":"https://openalex.org/keywords/data-modeling","display_name":"Data modeling","score":0.31369999051094055},{"id":"https://openalex.org/keywords/projection","display_name":"Projection (relational algebra)","score":0.30820000171661377}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7752000093460083},{"id":"https://openalex.org/C132010649","wikidata":"https://www.wikidata.org/wiki/Q189222","display_name":"Intuition","level":2,"score":0.5523999929428101},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.47920000553131104},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.4668000042438507},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.4578000009059906},{"id":"https://openalex.org/C51632099","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Training set","level":2,"score":0.3959999978542328},{"id":"https://openalex.org/C124101348","wikidata":"https://www.wikidata.org/wiki/Q172491","display_name":"Data mining","level":1,"score":0.3864000141620636},{"id":"https://openalex.org/C99498987","wikidata":"https://www.wikidata.org/wiki/Q2210247","display_name":"Noise (video)","level":3,"score":0.3831000030040741},{"id":"https://openalex.org/C2776145971","wikidata":"https://www.wikidata.org/wiki/Q30673951","display_name":"Labeled data","level":2,"score":0.3806000053882599},{"id":"https://openalex.org/C2776760102","wikidata":"https://www.wikidata.org/wiki/Q5139990","display_name":"Code (set theory)","level":3,"score":0.357699990272522},{"id":"https://openalex.org/C67186912","wikidata":"https://www.wikidata.org/wiki/Q367664","display_name":"Data modeling","level":2,"score":0.31369999051094055},{"id":"https://openalex.org/C57493831","wikidata":"https://www.wikidata.org/wiki/Q3134666","display_name":"Projection (relational algebra)","level":2,"score":0.30820000171661377},{"id":"https://openalex.org/C2781460075","wikidata":"https://www.wikidata.org/wiki/Q1399332","display_name":"Compliance (psychology)","level":2,"score":0.30079999566078186},{"id":"https://openalex.org/C63479239","wikidata":"https://www.wikidata.org/wiki/Q7353546","display_name":"Robustness (evolution)","level":3,"score":0.2994000017642975},{"id":"https://openalex.org/C2781170535","wikidata":"https://www.wikidata.org/wiki/Q30587856","display_name":"Noisy data","level":2,"score":0.289000004529953},{"id":"https://openalex.org/C55037315","wikidata":"https://www.wikidata.org/wiki/Q5421151","display_name":"Experimental data","level":2,"score":0.2851000130176544},{"id":"https://openalex.org/C133462117","wikidata":"https://www.wikidata.org/wiki/Q4929239","display_name":"Data collection","level":2,"score":0.28130000829696655},{"id":"https://openalex.org/C163985040","wikidata":"https://www.wikidata.org/wiki/Q1172399","display_name":"Data acquisition","level":2,"score":0.26260000467300415},{"id":"https://openalex.org/C43126263","wikidata":"https://www.wikidata.org/wiki/Q128751","display_name":"Source code","level":2,"score":0.2623000144958496},{"id":"https://openalex.org/C198531522","wikidata":"https://www.wikidata.org/wiki/Q485146","display_name":"Sample (material)","level":2,"score":0.257099986076355}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2606.00160","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.00160","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.2606.00160","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.00160","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":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Large":[0],"language":[1],"models":[2],"(LLMs)":[3],"suffer":[4,25],"from":[5,26],"degraded":[6],"safety":[7,83,171],"capabilities":[8],"even":[9],"when":[10],"fine-tuned":[11],"with":[12],"benign":[13,23,57,146],"datasets.":[14],"However,":[15],"existing":[16],"methods":[17],"for":[18],"identifying":[19,153],"safety-degrading":[20,46],"samples":[21],"in":[22,152],"datasets":[24,147],"high":[27],"computational":[28],"costs":[29],"and":[30,42,116,139,144,155,173],"significant":[31],"noise":[32],"issues.":[33],"In":[34],"this":[35,182],"paper,":[36],"we":[37],"propose":[38],"DataShield":[39,86],"to":[40,71,76,169,177],"efficiently":[41],"effectively":[43],"identify":[44],"potential":[45],"samples.":[47],"Our":[48],"key":[49,67],"intuition":[50],"is":[51,70,166,195],"based":[52],"on":[53,136],"the":[54,60,77,98,112,123,130,142],"observation":[55],"that":[56,162],"fine-tuning":[58],"increases":[59],"overall":[61],"response":[62],"compliance":[63,79,100,131],"of":[64,88,126],"LLMs.":[65],"DataShield's":[66],"technical":[68],"insight":[69],"quantify":[72],"each":[73],"sample's":[74],"contribution":[75],"model's":[78],"behavior":[80,101],"as":[81],"its":[82],"degradation":[84],"score.":[85],"consists":[87],"three":[89],"core":[90],"components:":[91],"(1)":[92],"Compliance":[93],"Vector":[94],"Extraction,":[95],"which":[96,109,121],"captures":[97],"LLM's":[99],"tendency;":[102],"(2)":[103],"a":[104],"novel":[105],"Compliance-Aware":[106],"Score":[107],"(CAS),":[108],"automatically":[110],"identifies":[111],"optimal":[113],"safety-critical":[114],"layer;":[115],"(3)":[117],"Safety-degrading":[118],"Sample":[119],"Filtering,":[120],"quantifies":[122],"projection":[124],"shift":[125],"training":[127],"data":[128,157],"along":[129],"direction.":[132],"Extensive":[133],"experimental":[134],"evaluation":[135],"Llama3-8B,":[137],"Llama3.1-8B,":[138],"Qwen2.5-7B":[140],"using":[141],"Alpaca":[143],"Dolly":[145],"validates":[148],"our":[149],"method's":[150],"effectiveness":[151],"high-risk":[154],"low-risk":[156],"subsets.":[158],"We":[159,180],"also":[160],"observe":[161],"open-ended":[163],"question":[164],"answering":[165],"more":[167],"likely":[168],"trigger":[170],"degradation,":[172],"corresponding":[174],"responses":[175],"tend":[176],"be":[178],"longer.":[179],"hope":[181],"work":[183],"can":[184],"provide":[185],"new":[186],"insights":[187],"into":[188],"data-centric":[189],"defense":[190],"methods.":[191],"The":[192],"source":[193],"code":[194],"available":[196],"at:":[197],"https://github.com/ZJunBo/DataShield.":[198]},"counts_by_year":[],"updated_date":"2026-07-01T06:00:48.157686","created_date":"2026-06-03T00:00:00"}