{"id":"https://openalex.org/W7160677126","doi":null,"title":"Evaluating Prompt Defenses on Small On-Device LLMs","display_name":"Evaluating Prompt Defenses on Small On-Device LLMs","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W7160677126"},"language":"en","primary_location":{"id":"pmh:oai:www.doria.fi:10024/194655","doi":null,"is_oa":true,"landing_page_url":"https://www.doria.fi/handle/10024/194655","pdf_url":"https://www.doria.fi/bitstream/10024/194655/2/lagerqvist_rasmus.pdf","source":{"id":"https://openalex.org/S4377196760","display_name":"Doria (University of Helsinki)","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I133731052","host_organization_name":"University of Helsinki","host_organization_lineage":["https://openalex.org/I133731052"],"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":"Diplomarbete"},"volume":null,"type":"other","is_traditional":false,"indexed_in":[],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://www.doria.fi/bitstream/10024/194655/2/lagerqvist_rasmus.pdf","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5135648617","display_name":"Rasmus Lagerqvist","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":true,"raw_author_name":"Lagerqvist, Rasmus","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":0,"corresponding_author_ids":["https://openalex.org/A5135648617"],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":true,"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":0.9994999766349792,"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":0.9994999766349792,"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/T11424","display_name":"Security and Verification in Computing","score":0.00019999999494757503,"subfield":{"id":"https://openalex.org/subfields/1708","display_name":"Hardware and Architecture"},"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/small-language-models","display_name":"small language models","score":0.8370000123977661},{"id":"https://openalex.org/keywords/qwen3-4b","display_name":"Qwen3-4B","score":0.8650000095367432},{"id":"https://openalex.org/keywords/llama-3-2-3b","display_name":"Llama-3.2-3B","score":0.9739999771118164},{"id":"https://openalex.org/keywords/attack-success-rate","display_name":"attack success rate","score":0.8100000023841858},{"id":"https://openalex.org/keywords/llm-as-a-judge","display_name":"LLM-as-a-judge","score":0.5249999761581421},{"id":"https://openalex.org/keywords/evaluation-uncertainty","display_name":"uncertainty in evaluation","score":0.5709999799728394},{"id":"https://openalex.org/keywords/prompt-injection","display_name":"prompt injection","score":0.31299999356269836}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6129000186920166},{"id":"https://openalex.org/C63479239","wikidata":"https://www.wikidata.org/wiki/Q7353546","display_name":"Robustness (evolution)","level":3,"score":0.5849999785423279},{"id":"https://openalex.org/C38652104","wikidata":"https://www.wikidata.org/wiki/Q3510521","display_name":"Computer security","level":1,"score":0.5115000009536743},{"id":"https://openalex.org/C177212765","wikidata":"https://www.wikidata.org/wiki/Q627335","display_name":"Workflow","level":2,"score":0.5115000009536743},{"id":"https://openalex.org/C199521495","wikidata":"https://www.wikidata.org/wiki/Q181487","display_name":"Audit","level":2,"score":0.490200012922287},{"id":"https://openalex.org/C55282118","wikidata":"https://www.wikidata.org/wiki/Q252683","display_name":"Snapshot (computer storage)","level":2,"score":0.4512999951839447},{"id":"https://openalex.org/C100660578","wikidata":"https://www.wikidata.org/wiki/Q18733","display_name":"Recall","level":2,"score":0.4366999864578247},{"id":"https://openalex.org/C124101348","wikidata":"https://www.wikidata.org/wiki/Q172491","display_name":"Data mining","level":1,"score":0.3156999945640564},{"id":"https://openalex.org/C197947376","wikidata":"https://www.wikidata.org/wiki/Q5155608","display_name":"Comparability","level":2,"score":0.3142000138759613},{"id":"https://openalex.org/C80958533","wikidata":"https://www.wikidata.org/wiki/Q1047174","display_name":"Audit trail","level":3,"score":0.31310001015663147},{"id":"https://openalex.org/C112930515","wikidata":"https://www.wikidata.org/wiki/Q4389547","display_name":"Risk analysis (engineering)","level":1,"score":0.3125999867916107},{"id":"https://openalex.org/C43214815","wikidata":"https://www.wikidata.org/wiki/Q7310987","display_name":"Reliability (semiconductor)","level":3,"score":0.3100000023841858},{"id":"https://openalex.org/C2780233690","wikidata":"https://www.wikidata.org/wiki/Q535347","display_name":"Transparency (behavior)","level":2,"score":0.2971999943256378},{"id":"https://openalex.org/C71745522","wikidata":"https://www.wikidata.org/wiki/Q2476929","display_name":"Confidentiality","level":2,"score":0.28940001130104065},{"id":"https://openalex.org/C82876162","wikidata":"https://www.wikidata.org/wiki/Q17096504","display_name":"Latency (audio)","level":2,"score":0.28859999775886536},{"id":"https://openalex.org/C2778024590","wikidata":"https://www.wikidata.org/wiki/Q2357432","display_name":"Cheating","level":2,"score":0.287200003862381},{"id":"https://openalex.org/C2776760102","wikidata":"https://www.wikidata.org/wiki/Q5139990","display_name":"Code (set theory)","level":3,"score":0.26010000705718994},{"id":"https://openalex.org/C200601418","wikidata":"https://www.wikidata.org/wiki/Q2193887","display_name":"Reliability engineering","level":1,"score":0.25859999656677246},{"id":"https://openalex.org/C2776889888","wikidata":"https://www.wikidata.org/wiki/Q1135789","display_name":"Unintended consequences","level":2,"score":0.2524000108242035},{"id":"https://openalex.org/C2987098735","wikidata":"https://www.wikidata.org/wiki/Q3808900","display_name":"Recall rate","level":2,"score":0.2522999942302704}],"mesh":[],"locations_count":1,"locations":[{"id":"pmh:oai:www.doria.fi:10024/194655","doi":null,"is_oa":true,"landing_page_url":"https://www.doria.fi/handle/10024/194655","pdf_url":"https://www.doria.fi/bitstream/10024/194655/2/lagerqvist_rasmus.pdf","source":{"id":"https://openalex.org/S4377196760","display_name":"Doria (University of Helsinki)","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I133731052","host_organization_name":"University of Helsinki","host_organization_lineage":["https://openalex.org/I133731052"],"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":"Diplomarbete"}],"best_oa_location":{"id":"pmh:oai:www.doria.fi:10024/194655","doi":null,"is_oa":true,"landing_page_url":"https://www.doria.fi/handle/10024/194655","pdf_url":"https://www.doria.fi/bitstream/10024/194655/2/lagerqvist_rasmus.pdf","source":{"id":"https://openalex.org/S4377196760","display_name":"Doria (University of Helsinki)","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I133731052","host_organization_name":"University of Helsinki","host_organization_lineage":["https://openalex.org/I133731052"],"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":"Diplomarbete"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7160677126.pdf","grobid_xml":"https://content.openalex.org/works/W7160677126.grobid-xml"},"referenced_works_count":12,"referenced_works":["https://openalex.org/W2020018978","https://openalex.org/W2626778328","https://openalex.org/W3030163527","https://openalex.org/W3133702157","https://openalex.org/W4311991106","https://openalex.org/W4385374425","https://openalex.org/W4385894687","https://openalex.org/W4388482289","https://openalex.org/W4388886073","https://openalex.org/W4389713766","https://openalex.org/W4391632220","https://openalex.org/W4415797606"],"related_works":["https://openalex.org/W4417085119","https://openalex.org/W7204164097","https://openalex.org/W7160692812","https://openalex.org/W4399554713","https://openalex.org/W7214543928","https://openalex.org/W7214347141","https://openalex.org/W7136611768","https://openalex.org/W7208759636","https://openalex.org/W7203780712","https://openalex.org/W4415844496"],"abstract_inverted_index":{"Small,":[0],"on-device":[1,241],"language":[2],"models":[3,50],"are":[4,117,135,259],"now":[5],"used":[6],"in":[7,42,57,261],"products":[8],"where":[9],"privacy":[10],"and":[11,14,32,62,76,80,87,98,137,152,167,193,225,234,257],"latency":[12],"matter,":[13],"the":[15,49,158,206,255,262],"safety":[16,159,232],"stack":[17],"varies":[18],"by":[19,119],"deployment.":[20],"This":[21],"thesis":[22,218,263],"measures":[23],"jailbreak":[24],"robustness":[25],"of":[26,201,223],"two":[27,77,94],"compact,":[28],"locally":[29],"run":[30,184],"models\u2014Qwen3-4B":[31],"Llama":[33,155,170],"3.2-3B\u2014in":[34],"single-turn,":[35],"text-only":[36],"settings.":[37],"The":[38,66,142,217,243],"evaluation":[39],"is":[40,91,148,209,246],"implemented":[41],"lfuzz,":[43],"a":[44,58,69,99,104,109,120,129,230,237],"local":[45,130],"framework":[46],"that":[47,102,112,123],"runs":[48],"through":[51],"Ollama,":[52],"wraps":[53],"each":[54,179],"harmful":[55,251],"goal":[56,90],"manual":[59,199],"attack":[60],"template,":[61],"applies":[63],"prompt-only":[64],"defenses.":[65],"experiments":[67],"use":[68],"frozen":[70],"100-goal":[71],"snapshot":[72],"from":[73],"JailbreakBench":[74],"Behaviors":[75],"seeds":[78],"(42":[79],"43),":[81],"yielding":[82],"200":[83],"evaluations":[84],"per":[85],"model":[86],"defense.":[88],"Each":[89],"evaluated":[92],"under":[93],"conditions:":[95],"no":[96],"defense":[97],"\u201csafety":[100],"sandwich\u201d":[101],"combines":[103,124],"real":[105],"system":[106],"policy":[107],"with":[108,128,172],"user-message":[110],"wrapper":[111],"marks":[113],"untrusted":[114],"text.":[115],"Responses":[116],"labeled":[118],"hybrid":[121],"evaluator":[122],"refusal":[125],"string":[126],"matching":[127],"LLM":[131],"judge;":[132],"uncertain":[133],"cases":[134],"flagged":[136],"not":[138],"counted":[139],"as":[140],"jailbreaks.":[141],"baseline":[143],"Attack":[144],"Success":[145],"Rate":[146],"(ASR)":[147],"25.0%":[149],"for":[150,154,165,169,178,187,196,220,240,264],"Qwen3-4B":[151,166],"38.0%":[153],"3.2-3B.":[156],"With":[157],"sandwich,":[160],"ASR":[161],"drops":[162],"to":[163],"0.5%":[164],"23.5%":[168],"3.2-3B,":[171],"Wilson":[173],"95%":[174],"confidence":[175],"intervals":[176],"reported":[177],"estimate.":[180],"A":[181,198],"matched-token":[182],"sensitivity":[183],"(4096":[185],"tokens":[186],"both":[188],"models)":[189],"preserves":[190],"these":[191],"conclusions":[192],"removes":[194],"truncation":[195],"Llama.":[197],"audit":[200],"400":[202],"seed-42":[203],"outputs":[204],"shows":[205],"automated":[207],"judge":[208,226],"high-precision":[210],"but":[211,254],"moderate-recall":[212],"(precision":[213],"0.93,":[214],"recall":[215],"0.74).":[216],"argues":[219],"transparent":[221],"reporting":[222],"uncertainty":[224],"reliability":[227],"rather":[228],"than":[229],"single":[231],"score,":[233],"it":[235,249],"provides":[236],"reproducible":[238],"workflow":[239],"evaluation.":[242],"code":[244],"repository":[245],"private":[247],"because":[248],"contains":[250],"prompt":[252],"material,":[253],"method":[256],"artifacts":[258],"documented":[260],"verification.":[265]},"counts_by_year":[],"updated_date":"2026-11-07T11:25:07.637180","created_date":"2026-05-09T00:00:00"}
