Federated learning на собеседовании Data Scientist

Идея federated learning

Train model на data, которая никогда не покидает client device.

Server → send model к clients
Clients → train на local data → send только updates
Server → aggregate updates → updated model

Data privacy preserved.

FedAvg

Standard algorithm (Google 2017).

1. Server initializes model w_0.
2. Каждый round:
   a. Send w_t to N selected clients.
   b. Каждый client trains несколько epochs локально → получает w_i_t+1.
   c. Send updates back.
   d. Server averages: w_t+1 = sum(n_i * w_i_t+1) / sum(n_i).

n_i — sample count client i.

Variations: FedProx (regularization), FedNova, SCAFFOLD.

Privacy techniques

Secure aggregation. Cryptographic protocol — server видит aggregate, не individual updates.

Differential privacy. Add noise to gradients перед send. Bounded info leak per client.

Homomorphic encryption. Compute на encrypted data. Powerful, dorogo.

Cross-device vs cross-silo

Cross-device. Millions устройств (mobile phones, wearables). Each has small data.

• Examples: Google Gboard, Apple Siri.
• Challenges: massive scale, intermittent connectivity, device heterogeneity.

Cross-silo. Несколько organizations (hospitals, banks). Each имеет large data.

• Examples: medical research across hospitals.
• Challenges: trust, governance, IID violation.

Применения

Mobile keyboards. Personalized predictions без uploading typing history.

Medical research. Multiple hospitals collaborate без sharing patients data.

Finance. Banks fraud detection cross-bank без sharing customer data.

IoT. Devices learn без sending raw sensor data.

В РФ: federated learning редкость в production. Иногда research projects.

Связанные темы

• Privacy и 152-ФЗ для SA
• MLOps на собесе DS
• Gradient descent для DS
• Bias и fairness для DS
• Подготовка к собесу Data Scientist

FAQ

Это официальная информация?

Нет. Статья основана на работах McMahan 2017 (FedAvg), документации TFF / Flower.

Тренируйте Data Science — откройте тренажёр с 1500+ вопросами для собесов.