With　the　widespread　adoption　of　biometric　technologies,　data　privacy　and　security　have　become　pressing　concerns.　This　study　proposes　a　three-party　iris　recognition　model　based　on　fully　homomorphic　encryption　(FHE)　and　an　improved　center-symmetric　local　binary　pattern　(CS-LBP)　feature　extraction　algorithm　to　address　these　challenges.　The　system　comprises　a　client,　server,　and　third-party　computation　entity.　The　client　first　extracts　a　256-dimensional　feature　vector　from　the　iris　image　using　the　enhanced　CS-LBP　algorithm,　improving　the　accuracy　and　robustness　of　feature　representation.　The　feature　vector　is　then　encrypted　using　the　Brakerski-Gentry-Vaikuntanathan　(BGV)　homomorphic　encryption　scheme　and　transmitted　to　the　third-party　computation　entity.　Given　the　computational　demands　of　the　BGV　algorithm　and　the　256-dimensional　feature　vector,　the　inclusion　of　cloud　computing　significantly　reduces　the　impact　of　these　demands,　enabling　efficient　feature　matching　and　similarity　calculations　in　the　encrypted　domain.　This　approach　effectively　prevents　data　breaches　and　misuse,　safeguarding　user　privacy.　Experiments　conducted　on　multiple　subsets　of　the　CASIA-IrisV4　dataset　under　various　conditions　show　that　the　proposed　model　excels　across　key　performance　metrics,　with　a　recognition　accuracy　of　97.0%,　a　recall　rate　of　96.5%,　an　F1　score　of　96.7%,　and　an　area　under　the　curve　(AUC)　of　0.990.　While　the　system　faces　challenges　in　time　efficiency,　cloud　computing　integration　helps　mitigate　these,　resulting　in　an　average　feature　extraction　time　of　0.6　seconds　per　image,　encryption　time　of　0.3　seconds　per　feature　vector,　and　homomorphic　computation　time　of　2.0　seconds　per　match.　This　study　demonstrates　the　system＇s　superiority　in　maintaining　recognition　accuracy　and　improving　computational　efficiency　while　protecting　user　privacy.　Beyond　its　potential　in　iris　recognition,　it　offers　a　novel　solution　for　privacy　protection　in　other　biometric　applications.　Future　research　will　focus　on　further　optimizing　the　computational　efficiency　of　the　encryption　algorithm,　expanding　the　model＇s　applicability　to　different　biometric　domains,　and　exploring　efficient　implementations　in　cloud　computing　environments.　©　2025　SPIE.