FEDERATIV O‘QITISH ASOSIDA MAXFIY MA’LUMOTLAR BILAN ISHLASHDA SUN’IY INTELLEKT MODELLARINI XAVFSIZ HAMKORLIKDA O‘QITISH METODOLOGIYASI

Abstract

Sun’iy intellekt tizimlarini yuqori sifatli ma’lumotlar asosida o‘qitish ko‘pincha ma’lumotlarni markazlashtirishni talab qiladi. Biroq tibbiyot, moliya, davlat boshqaruvi va telekommunikatsiya kabi sohalarda ma’lumotlar maxfiyligi, huquqiy cheklovlar va tashkilotlararo ishonch masalalari ularni yagona bazaga yig‘ishni qiyinlashtiradi. Ushbu maqolada xom ma’lumotlarni tashkilotlar tashqarisiga chiqarmasdan birgalikda model o‘qitishga imkon beruvchi federativ o‘qitish (federated learning, FL) uchun xavfsizlik va boshqaruvga yo‘naltirilgan metodologik arxitektura taklif qilinadi. Yondashuv mahalliy o‘qitish, vaznlarni agregatsiya qilish, xavfsiz agregatsiya, differensial maxfiylik, ma’lumotlar taqsimotining notekisligi, model sifati va auditni yagona jarayonda ko‘rib chiqadi. Maqola mavjud ilmiy ishlar sinteziga asoslanadi va yangi empirik tajriba o‘tkazilganini da’vo qilmaydi. Uning asosiy hissasi federativ o‘qitish loyihalarini rejalashtirish uchun bosqichma-bosqich protokol, xavf modeli va baholash mezonlarini tizimlashtirishdir. Taklif etilgan metodologiya maxfiylikni saqlashni mutlaq kafolat deb emas, balki texnik, tashkiliy va statistik nazoratlar bilan boshqarilishi lozim bo‘lgan xavf sifatida talqin qiladi.

Training artificial intelligence systems often requires access to large and diverse datasets. In healthcare, finance, public administration, and telecommunications, however, privacy obligations, legal restrictions, and inter-organizational trust can make centralizing raw data impractical. This article proposes a security- and governance-oriented methodological architecture for federated learning (FL), which enables collaborative model training while keeping raw data at participating organizations. The framework integrates local training, model aggregation, secure aggregation, differential privacy, non-identically distributed data, model utility evaluation, and auditability. It synthesizes established research and does not claim new empirical experiments. Its contribution is a structured protocol, threat model, and evaluation criteria for planning federated learning projects. Privacy is treated not as an automatic guarantee of FL, but as a risk that must be managed through technical, organizational, and statistical controls.