A form of distributed machine learning known as “federated learning” uses data from edge devices, such as laptops, smartphones, and wearable technology, to train machine learning and deep learning algorithms without transferring the data to a central server.
Among the several advantages it confers are meeting latency constraints, promoting data privacy and security, and making parameter updates in a distributed manner.
It is thus a decentralized approach to machine learning, where, across multiple organizations or devices, data can be used to collaboratively build machine learning models without anyone sharing the actual private data. Instead of raw data being moved to some central server, only some updates or parameter values are exchanged, thus ensuring the privacy of the data and also its security.
Federated learning is an approach that thereby supports data privacy on the one hand, in that training data remains local and only aggregated insights are exchanged, while on the other hand, the federated data are used for improving model accuracy.
Types of Federated Learning:
- Horizontal Federated Learning:
Horizontal federated learning protects privacy by allowing several parties with distinct users but comparable data attributes to work together to build a model without exchanging raw data.
- Vertical Federated Learning
Vertical Federated Learning occurs when multiple clients share the same users but possess different features. It enables collaborative model training across organizations that hold complementary data about the same individuals, without exchanging raw data.
- Federated Transfer Learning:
Federated transfer learning is basically making federated learning meet transfer learning so that clients with different data can collaborate. This allows models to transfer knowledge even if the clients have different features and user distributions, thus aiding a common project in optimizing its performance without the exchange of raw data.
Federated learning can also be divided into two categories based on the size of the participating clients: Cross-Device Federated Learning and Cross-Silo Federated Learning.
How federated learning works:
Federated learning is a privacy-preserving machine learning technology by which multiple devices or organizations set about building a shared model collaboratively without disclosing any raw data. A central server starts the process by selecting a global model and disseminating it among their client devices. Each client trains the model on their own private dataset, hence sensitive information remains on the client device. When training is completed, clients submit only the updated model parameters (weights or gradients) to the server. The aggregator then combines the clients’ updates, usually performing an averaging operation known as Federated Averaging or FedAvg, with to update the global model. This improved global model is redistributed for more rounds of training, and so on. Consequently, the model learns from several data sources while ensuring the privacy and security of the data. This is especially useful in hospitals, finances, and mobile apps.
Applications of Federated Learning:
- Autonomous Vehicle:
Federated learning enables self-driving cars to be safer and smarter through real-time awareness of road terrain, faster decisions on the spot, and continuous model updating. Vehicles share insights locally like hazards or weather changes without sending raw data, allowing onboard AI to react instantly while improving overall system accuracy over time.
- Mobile and Edge Devices
FL enables more intelligent and private user experiences in mobile technologies. For instance, Google Gboard learns from user typing behaviors right on the device to enhance text predictions. Through local training, voice assistants such as Google Assistant and Siri improve speech recognition and customisation. Without jeopardizing user privacy, FL also offers individualized content recommendations.
- Industrial IoT
Federated learning in IoT allows machines and sensors to train models locally with their own data while not actually sharing it. Only these model updates are communicated to the central server for a combined update in performance. This serves predictive maintenance and anomaly detection while rendering operational data private and secure.
- Finance
In the financial industry, FL is a method for banks and other financial institutions to collaborate against fraud, measuring creditworthiness, evaluating market risks, and so forth. Training of the model occurs on distributed data sources, thus providing institutions with a wider perspective while preserving customer building laws with regards to data sovereignty.
- Cybersecurity
With the FL approach, one can detect anomalies and forecast malicious threats on the basis of observed local attack patterns. This constitutes a decentralized approach toward developing defense, thereby ensuring that sensitive logs are not merged together. Biometric authentication systems take one step further by endowing local training that keeps personal identifiers locked on the device.
Federated learning advantages:
Federating learning keeps data in a local device, which enhances privacy and security. It reduces bandwidth usage, supports personalized models, and allows learning from a broader and diversified data source without centralizing information that is disclosive.
Federated learning disadvantages:
It requires high resources on the device, lacks consistency in data distribution over users, and faces barriers of coordination and debugging. It may also train slow the models and be less accurate than centralized ones.
Federated Learning Examples:
Some of the use cases for Federated learning are:
- Google Gboard: Improve predictive text and suggestions without the need to upload data of user typing.
- Healthcare: Hospitals train the model on patient data-essentially-training without sharing sensitive records.
- Finance: Banks employ federated learning to detect fraud across institutions without exposing customer data.
- Google: Google uses FL to enhance on-device machine learning systems, such as the “Hey Google” detection in Google Assistant, enabling users to issue voice commands.
