1. What is the main goal of federated learning?

The main goal is to train machine learning models across multiple devices or organizations without requiring raw data to leave those devices. Each participant trains the model locally, and only model updates are shared with a central server. This keeps sensitive data private while still allowing a global model to improve from diverse data.

2. Who invented federated learning?

Federated learning was introduced by researchers at Google in 2016 and formally described in a 2017 paper titled "Communication-Efficient Learning of Deep Networks from Decentralized Data" by McMahan et al. The original use case was improving the Gboard keyboard on Android devices without accessing user typing data on Google's servers.

3. Is federated learning the same as distributed learning?

No. Distributed learning splits data across machines to speed up training, typically in a controlled data center where all nodes are trusted and data can move freely. Federated learning specifically addresses the case where data cannot or should not leave local devices, and participants may be untrusted or unreliable. Privacy is the key distinction.

4. How is model accuracy affected in federated learning?

Accuracy can be lower than centralized training, especially when data across devices is non-IID. When one hospital sees mostly one type of patient data, the local model may overfit to that distribution. Algorithms like FedProx and SCAFFOLD help reduce this accuracy gap, but centralized training with full data access generally still outperforms federated approaches in controlled experiments.

FedAvg, short for Federated Averaging, is the baseline aggregation algorithm for federated learning. After each round of local training, participating devices send their updated model weights to the central server. The server computes a weighted average of those weights, where the weight given to each device is proportional to the size of its local dataset. This averaged model becomes the new global model.

6. Can federated learning completely prevent data leaks?

Federated learning reduces the risk significantly but does not eliminate it entirely. Gradient inversion attacks have demonstrated that an attacker with access to gradient updates can sometimes reconstruct parts of the original training data. To strengthen privacy guarantees, federated learning is often combined with differential privacy, which adds controlled noise to the updates before they are shared.

7. What is the role of differential privacy in federated learning?

Differential privacy adds calibrated statistical noise to the model updates before they are sent to the server. This makes it mathematically difficult for anyone to infer whether a specific data point was part of a device's local training set. The tradeoff is that adding noise can reduce model accuracy.

8. What industries are adopting federated learning most aggressively in 2026?

Healthcare and finance are the leading sectors. Hospitals use it to train diagnostic models across institutions without sharing patient records. Banks use it for fraud detection and credit risk modeling without exposing transaction data. Telecom companies use it for network optimization.

9. What is cross-silo versus cross-device federated learning?

Cross-device federated learning involves a large number of consumer devices like smartphones, often millions, with small amounts of data each. Cross-silo federated learning involves a small number of organizations, such as hospitals or banks, each holding large, well-curated datasets.

10. How do you get started with federated learning practically?

The easiest starting point is the Flower (flwr) framework in Python. It works with PyTorch and TensorFlow and lets you simulate a federated setup locally on your machine before deploying to real distributed nodes. The official documentation includes tutorials on image classification and NLP tasks.

11. Will federated learning work with large language models?

Yes, but it is challenging. LLMs have billions of parameters, and even sharing gradient updates is costly in terms of communication bandwidth. Researchers are actively working on federated fine-tuning approaches using parameter-efficient methods like LoRA, which only update a small subset of parameters.