Introduction to Federated Learning: Privacy-Preserving AI - Dr Kokilavani T

---

The invention of AI has significantly impacted human life. The popularity of machine learning, one of the AI technologies, has increased recently because of its usage in business and research. To make machines think and work smarter like humans, they must be trained with large amounts of real data. Data is most valuable for any organization or individual. Organizations use data for decision-making, to improve productivity, or to optimize their operations. Data privacy is important for individuals, and security has become a major concern for many organizations when providing their data. In traditional machine learning, the data is stored in a centralized server to train the model. This leads to the possibility of a data leak. 

Federated Learning (FL) is a type of Machine Learning (ML) technology that follows a decentralized approach in training models, whereas traditional machine learning follows a centralized approach. The concept of FL was first proposed by Google in 2016 and mainly focuses on Android mobile phone users to protect their private personal data. Federated Learning is a distributed approach involving several devices that train a model collaboratively without sharing their personal data. Each device trains the model locally and sends the updates to a central server responsible for aggregating the results to the shared model. In this approach, data privacy is preserved because each device used for training shares only the information necessary to update the model while maintaining its data locally. Federated Learning, also known as collaborative learning, can be applied to multiple industrial applications, but each domain has its own set of risks. FL can be used for various applications, including defence, healthcare, transportation, IoT, natural language processing, and mobile apps. Medical data always includes highly sensitive patient information, which should be protected from unauthorized access.

Two types of architectures are followed for FL. They are: Horizontal FL and Vertical FL. Both are used for different types of applications. In horizontal FL, the data set is similar but users are different, whereas in vertical FL, the data set is different but the users are similar. Some of the machine learning techniques, like linear regression, classification, gradient descent, and association rule mining, can use vertical FL, whereas deep learning techniques can be applied for horizontal FL. Wireless networks and Edge computing are used to implement Federated Learning. 

The efficiency of federated learning mainly depends on the aggregation of local model updates, which leads to the formation of a global model combining the knowledge gathered from all participating nodes. At the same time, the model should also consider the heterogeneity of participating nodes, communication overhead occurring because of geographical distribution, and privacy-preserving issues.

Security and privacy-preserving are two important features of Federated Learning. Some of the traditional encryption techniques like homomorphic encryption, differential privacy, and multi-party computing applied to machine learning can also be used for federated Learning. Since federated learning maintains data privacy with the concept of decentralized data, it is getting attention among academicians and industrialists.