{"id":"https://openalex.org/W7166802823","doi":"https://doi.org/10.18653/v1/2026.acl-long.1310","title":"OASIS: Mitigating Harmful Fine-tuning Attacks on LLMs via Orthogonal and Adaptive Safety Alignment Strategy","display_name":"OASIS: Mitigating Harmful Fine-tuning Attacks on LLMs via Orthogonal and Adaptive Safety Alignment Strategy","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W7166802823","doi":"https://doi.org/10.18653/v1/2026.acl-long.1310"},"language":null,"primary_location":{"id":"doi:10.18653/v1/2026.acl-long.1310","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.1310","pdf_url":"https://aclanthology.org/2026.acl-long.1310.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"},"type":"conference-paper","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://aclanthology.org/2026.acl-long.1310.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5113820036","display_name":"Jiayu Tang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Jiayu Tang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139852872","display_name":"Guowei Peng","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Guowei Peng","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5001625286","display_name":"Qiuhao Xie","orcid":"https://orcid.org/0000-0002-8389-6580"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Qiuhao Xie","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139844641","display_name":"Yuning Yang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yuning Yang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5135720503","display_name":"Xiurui Xie","orcid":"https://orcid.org/0000-0002-3720-4379"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Xiurui Xie","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5139837538","display_name":"Guisong Liu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Guisong Liu","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":0.0,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":{"value":0.79223854,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"28407","last_page":"28421"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T11424","display_name":"Security and Verification in Computing","score":0.18279999494552612,"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/T11424","display_name":"Security and Verification in Computing","score":0.18279999494552612,"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/T11689","display_name":"Adversarial Robustness in Machine Learning","score":0.13420000672340393,"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/T10400","display_name":"Network Security and Intrusion Detection","score":0.10140000283718109,"subfield":{"id":"https://openalex.org/subfields/1705","display_name":"Computer Networks and Communications"},"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/key","display_name":"Key (lock)","score":0.37709999084472656},{"id":"https://openalex.org/keywords/vulnerability","display_name":"Vulnerability (computing)","score":0.3343000113964081},{"id":"https://openalex.org/keywords/risk-management","display_name":"Risk management","score":0.2865999937057495},{"id":"https://openalex.org/keywords/risk-assessment","display_name":"Risk assessment","score":0.2628999948501587},{"id":"https://openalex.org/keywords/control","display_name":"Control (management)","score":0.2484000027179718}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5357999801635742},{"id":"https://openalex.org/C38652104","wikidata":"https://www.wikidata.org/wiki/Q3510521","display_name":"Computer security","level":1,"score":0.5202999711036682},{"id":"https://openalex.org/C112930515","wikidata":"https://www.wikidata.org/wiki/Q4389547","display_name":"Risk analysis (engineering)","level":1,"score":0.5120000243186951},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.37709999084472656},{"id":"https://openalex.org/C144133560","wikidata":"https://www.wikidata.org/wiki/Q4830453","display_name":"Business","level":0,"score":0.3716000020503998},{"id":"https://openalex.org/C95713431","wikidata":"https://www.wikidata.org/wiki/Q631425","display_name":"Vulnerability (computing)","level":2,"score":0.3343000113964081},{"id":"https://openalex.org/C32896092","wikidata":"https://www.wikidata.org/wiki/Q189447","display_name":"Risk management","level":2,"score":0.2865999937057495},{"id":"https://openalex.org/C12174686","wikidata":"https://www.wikidata.org/wiki/Q1058438","display_name":"Risk assessment","level":2,"score":0.2628999948501587},{"id":"https://openalex.org/C2775924081","wikidata":"https://www.wikidata.org/wiki/Q55608371","display_name":"Control (management)","level":2,"score":0.2484000027179718},{"id":"https://openalex.org/C2780791683","wikidata":"https://www.wikidata.org/wiki/Q846785","display_name":"Action (physics)","level":2,"score":0.24779999256134033}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2026.acl-long.1310","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.1310","pdf_url":"https://aclanthology.org/2026.acl-long.1310.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.18653/v1/2026.acl-long.1310","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.1310","pdf_url":"https://aclanthology.org/2026.acl-long.1310.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"},"sustainable_development_goals":[{"score":0.5719354152679443,"id":"https://metadata.un.org/sdg/13","display_name":"Climate action"}],"awards":[{"id":"https://openalex.org/G60673156","display_name":null,"funder_award_id":"62376228","funder_id":"https://openalex.org/F4320321001","funder_display_name":"National Natural Science Foundation of China"}],"funders":[{"id":"https://openalex.org/F4320321001","display_name":"National Natural Science Foundation of China","ror":"https://ror.org/01h0zpd94"},{"id":"https://openalex.org/F4320336736","display_name":"Chengdu Science and Technology Program","ror":null}],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7166802823.pdf","grobid_xml":"https://content.openalex.org/works/W7166802823.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"The":[0],"\"Fine-Tuning-as-a-Service\"":[1],"paradigm":[2],"exposes":[3],"large":[4],"language":[5],"models":[6],"to":[7,23,67,79,121],"catastrophic":[8],"safety":[9,70],"degradation":[10],"from":[11,72],"less":[12],"harmful":[13,30,42,73,80],"samples.Alignment-stage":[14],"defenses":[15],"address":[16,55],"this":[17,56],"by":[18,117],"proactively":[19],"injecting":[20],"adversarial":[21],"perturbations":[22,77],"bolster":[24],"the":[25,46,114],"model's":[26],"inherent":[27],"robustness":[28],"against":[29],"drift.However,":[31],"existing":[32],"methods":[33],"rely":[34],"on":[35,85,100],"perturbation":[36],"directions":[37],"that":[38,111],"often":[39],"conflict":[40],"with":[41],"gradients,":[43],"inadvertently":[44],"facilitating":[45],"acquisition":[47],"of":[48],"malicious":[49],"features":[50],"rather":[51],"than":[52],"suppressing":[53],"them.To":[54],"issue,":[57],"we":[58],"propose":[59],"Orthogonal":[60],"and":[61,82,108],"Adaptive":[62],"Safety":[63],"Alignment":[64],"Strategy":[65],"(OASIS)":[66],"mathematically":[68],"decouple":[69],"enforcement":[71],"feature":[74],"acquisition.By":[75],"projecting":[76],"orthogonal":[78],"gradients":[81],"concentrating":[83],"optimization":[84],"adaptively":[86],"selected":[87],"safetycritical":[88],"layers,":[89],"OASIS":[90,112],"effectively":[91],"resolves":[92],"directional":[93],"conflicts":[94],"while":[95,124],"maximizing":[96],"parameter":[97],"efficiency.Extensive":[98],"experiments":[99],"four":[101],"LLMs":[102],"across":[103],"three":[104],"datasets":[105],"(SST2,":[106],"GSM8K,":[107],"AGNews)":[109],"demonstrate":[110],"reduces":[113],"Harmful":[115],"Score":[116],"approximately":[118],"60%":[119],"compared":[120],"competitive":[122],"baselines,":[123],"maintaining":[125],"stable":[126],"downstream":[127],"task":[128],"utility.":[129]},"counts_by_year":[],"updated_date":"2026-07-29T14:22:42.915294","created_date":"2026-07-02T00:00:00"}