Selected Applications
As an innovative modeling mechanism that can build shared and personalized models on decentralized data scattered among multiple parties without compromising user privacy and security, federated learning has promising applications in many important fields such as sales, finance, healthcare, education, urban computing, edge computing, and blockchain, where data cannot be directly aggregated for training machine learning (ML) models due to various reasons. In this chapter, we give an overview of several ongoing and potential applications that could be realized with federated learning beyond the current horizon.
The financial industry is greatly affected by government regulations in many ways for protecting investors against mismanagement and fraud, sustaining the stability of the financial sector, preserving the privacy and security of user data, and many others. To save cost and workload from government regulations, many financial and banking companies have exploited modern technologies such as AI, cloud services, and mobile Internet technologies to efficiently and effectively provide financial services while complying with strict government regulations.
Take smart consumer financing as an example. The purpose is to leverage ML techniques to offer personalized financial services to creditworthy consumers to encourage consumption. The data features involved in consumer finance mainly include consumer qualification information, purchasing power, and purchasing preference, as well as product characteristics. In practical applications, these data features are likely to be collected by different departments or companies. For example (Figure 10.1), a consumer’s qualification information and purchasing power can be inferred from her bank savings, and her purchasing preference for various products or services can be analyzed from her social networks. The characteristics of products are recorded by an e-shop. In this scenario, we are faced with two problems. First, for the protection of user privacy and data security, data barriers between banks, social networking sites, and the e-shopping sites are difficult to break. As a result, the data cannot be directly aggregated. Second, the data stored by the three parties are usually heterogeneous, and traditional ML models cannot directly work on heterogeneous data. For now, these problems have not been efficiently solved with traditional ML methods.
Federated learning and transfer learning are the key to solving these problems. First, based on federated learning, we can build local personalized models for the three parties without exposing their data. Meanwhile, we can leverage transfer learning to address the data heterogeneity problem, and overcome the limitations of traditional AI techniques. Therefore, federated learning provides excellent technical support for us to build a cross-enterprise, cross-data and cross-domain ecosystem for big data and AI.
Figure 10.1: Federated learning in smart consumer finance.
With the advance in AI technologies, many AI applications have been developed in the medical field with the hope to reduce labor costs and human errors. For example, AI programs for cardiology and radiology have been developed to help diagnose heart diseases and identify cancer cells in the early stages. With the promising applications of health AI, more and more healthcare providers are leveraging AI to create efficiency and improve patient care (Figure 10.2).
However, the adoption of AI technologies in the medical industry remains in its infancy. Existing intelligent medical systems are far from really “intelligence,” and some are being questioned for offering unsafe and incorrect treatment recommendations [Chen, 2018, Mearian, 2018]. Many factors may contribute to the deficiency of existing intelligent medical systems. A crucial one is the difficulty in collecting a sufficiently large amount of data with rich features that can comprehensively describe the symptom of a patient. For example, to accurately diagnose a disease, we may need features from various sources including disease symptoms, gene sequences, medical reports, examination results, and academic papers. However, there is no stable data source for filling in values of all those features. Besides, the labels of the majority of the training data are missing. Researchers estimate that it would take 10 years with 10,000 experts to gather a dataset useful enough for developing healthcare AI. The insufficiency of data and labels that results in the poor performance of ML models becomes the bottleneck of intelligent medical systems.
Figure 10.2: Federated learning in intelligent diagnosis.
To break through this bottleneck, medical institutions could unite together by sharing their data in compliance with privacy protection regulations. Then, we could possess a dataset large enough to train a model that can perform much better than the model trained on data from a single medical institution. Combining federated learning with transfer learning is a promising solution to achieve this goal. First, data from medical institutions are sensitive to privacy and security issues. Directly gathering such data in one location is infeasible. Federated learning allows all participating parties to collaboratively train a shared model without exchanging or exposing their private patient data. Second, transfer learning techniques can help expand the sample and feature space of training data and, in turn, improve the performance of the shared model. Therefore, federated transfer learning can play an important role in the development of intelligent medical systems. If a decent amount of the medical institutions could establish a data alliance together through federated learning in the future, health AI can bring more benefits to more patients.
Educators have long called for instructional systems that integrate cross-curricular subjects (e.g., among science, technology, engineering, and mathematics (STEM) subjects and also between STEM and the humanities). However, instructional systems can seldom handle the prerequisite skills, knowledge bases, and experiences necessary to provide such an integrated learning experience. A typical adaptive instructional system (AIS) addresses a single subject at a time, and it often has a unique content ontology, adaptive engine, and data management method. For example, a maths AIS ontology typically consists of a knowledge graph of granular learning objectives in maths, but it can have many connections to objectives in physics and chemistry. A student’s calculus knowledge, for example, could inform their learning experience in physics or chemistry. Thus, an integration of ontologies across instructional systems would not only expand the scope of multiple AISs, but also support a richer, cross-curricular adaptive learning experience for students.
To this end, we can encode the knowledge graph of each AIS as a graph neural network proven to have high representational power. Then, we can use federated learning-based approaches to build a comprehensive model that integrates knowledge neural graphs of various AIS, thereby extending the curricular knowledge, the learner model, and the data reach from one AIS to another. In this way, each AIS benefits from data synchronization, latency reduction, and security features of a federated system while maintaining its own ontology, adaptive engine, and data.
In addition to integrating educational resources, federated learning can help achieve personalized education (Figure 10.3). More specifically, educational institutions can utilize federated learning to collaboratively build a general learning-plan formulating model based on data stored at students’ personal mobile devices such as smartphones, iPads, and laptops. The general model can formulate a standardized learning plan for students having similar backgrounds. On top of that general model, a personalized model that can provide personalized learning instructions can be built for each student based on that student’s strengths, needs, skills, and interests.
10.4 URBAN COMPUTING AND SMART CITY
According to Zheng et al. [2013], urban computing is defined as a process of acquiring, integrating, and analyzing big and heterogeneous data generated by a diversity of sources in urban spaces, such as sensors, devices, vehicles, buildings, and humans, for tackling the major issues that cities face, such as air pollution, increased energy consumption, and traffic congestion. It is the technology that helps create smart cities aiming to agilely respond to citizens’ needs.
With the development of cloud services, big data, AI, the Internet of Things (IoT) and fifth generation (5G) technologies, smart cities were being built at an increasingly fast pace in many developing and developed countries. After its transient eruption, however, the development of smart cities has entered a slower phase in which cities face many big challenges. In iPesearch [2019], four challenges that researchers, engineers, and civil servants have been encountered while building smart cities are summarized.
Figure 10.3: Federated learning in education.
• Emphasizing on technology while ignoring participation. Emphasizing on informatization and platform construction among large enterprises and institutions, while ignoring the participation of the majority of small businesses.
• Data silos and data fragmentation. The lack of integration of data, applications, and departmental responsibilities for urban management remains unresolved.
• Security risks of intelligent systems. Insufficient attention is being paid to information security, operational security, and network security, which increases city management costs and risks.
• Lack of sustainable model of operation. The market participation mechanism is not comprehensive. Sustainable and just payoff-sharing and reward mechanism regulated by market rules need to be built.
Federated learning with its collaborative and privacy-preserving nature is a promising solution for addressing these challenges (Figure 10.4). Federated learning brings greater opportunities and benefits to small businesses through the collaborative construction of smart city technologies. Under the federated learning, small businesses can collaboratively build intelligent applications by exploiting the data of all participants without compromising privacy and security. For example, by applying federated learning, ride-hailing companies can collaboratively build optimal models to address the vehicle routing problem, not only directly increasing their revenue and improving customer satisfaction, but also gaining side benefits brought by distributing and reducing the traffic congestion.
Figure 10.4: Federated learning in urban computing and smart city.
With federated learning, the issues of data silos can be addressed to some extent. There are many factors that may contribute to data silos, such as regulatory risk, privacy concerns, misaligned incentives, and the high cost of integrating heterogeneous data. In addition to addressing the privacy concerns and bringing benefits to participating parties, federated learning is capable of integrating data with heterogeneous features. For example, current air quality prediction models generally rely on Air Quality Index (AQI) readings from sparsely distributed air quality monitoring stations and meteorological conditions, and are unable to make use of more fine resolution industrial emissions and vehicle exhaust data that have quite different feature spaces from AQI. Vertical federated learning is able to solve this problem, allowing the training of a virtual shared model on data with heterogeneous features.
In order to sustain long-term stability in a data federation and attract more high quality data owners over time, an incentive scheme that shares the profit generated by a federation with participants in a fair and just manner is needed. Federate Learning Incentivizer (FLI) was proposed for this function. We refer interested reader to Chapter 7 for details. The core of this payoff-sharing scheme is to dynamically divides a given budget among data owners in a federation by jointly maximizing the sustainable operation objectives, while minimizing the inequity among the data owners. With FLI, we envisage that more and more data owners will be motivated to contribute high-quality data to the data federation to facilitate the development of smart cities.
10.5 EDGE COMPUTING AND INTERNET OF THINGS
With the soaring number of netizens [CNNIC, 2018, eMarketer, 2017], the popularity of the mobile Internet and mobile phones has promoted the development of Mobile Edge Computing (MEC). MEC allows computing to occur where data are produced (i.e., on the IoT devices) instead of sending data to cloud servers. It can be applied to any single enterprise or institution that has deployed IoT devices, especially mobile devices.
A variety of applications powered by AI techniques (e.g., face recognition, voice assistant, and intelligent background blur) can be deployed on mobile phones. Current solutions for AI applications typically require user data to be uploaded to the cloud server in order to train a giant model. However, this may cause privacy violation and security breach. In addition, with the centralized nature of current AI algorithms, users may suffer from high-latency while using AI apps, especially when connection is weak.
Federated learning allows for building more intelligent models while preserving privacy and security of local data. It can serve as an operating system for edge computing, as it provides the learning protocols for coordination and security (Figure 10.5). More specifically, federated learning enables edge computing devices to collaboratively train an ML model without sending data to the cloud. In addition to the privacy-preserving benefit brought by federated learning, each mobile device ends up with a personalized model that can respond to users immediately. Google has ready tested its federated learning in the Gboard application on Android smartphones [McMahan and Ramage, 2017]. Their federated learning algorithm utilizes clicking histories of user query suggestions stored on devices to make improvements to the next iteration of Gboard’s query suggestion model.
The changes brought about by federated learning are not limited to mobile devices, but smart home terminals as well. Federated learning can make full use of heterogeneous features of data collected by different home devices to build smarter applications. Federated learning models are not only suitable for TV and electric lights, but also can be combined with smart speakers and door lockers for linkage function development, supporting more complex new modes of operation in smart homes.
The bringing model training to edge with federated learning opens up many algorithmic and technical challenges. One of them is that it requires edge devices to be installed with more powerful processors for training complex local models. This requirement will push terminal device makers such as Apple, Huawei, and Xiaomi to develop specialized hardware (e.g., neural processing unit (NPU)) tailored to modern AI techniques such as deep neural networks. With the evolution of AI and IoT, AI technologies, and edge computing will not be developing in isolation, but will be moving toward the path of integrated development.
Figure 10.5: Federated learning in edge computing.
Federated learning provides participants with the capability of collaboratively building powerful ML models and employs privacy-preserving mechanisms to protect the privacy of their data. However, federated learning has been questioned for its vulnerability to backdoor attacks [Bagdasaryan et al., 2019]. For example, malicious participants can poison machines learning models with malicious training samples and undermine the effects of the final model using model replacement techniques. Many secure protocols have been developed to guard against malicious attacks (e.g., defensive distillation and adversarial training regularization). However, in order to actively prevent federated learning from malicious attacks instead of just passive defense, a mechanism that can effectively detect malicious attacks and pinpoint malicious participants is needed.
Blockchain, with its immutability and traceability, can be an effective tool to prevent malicious attacks in federated learning [Preuveneers et al., 2018]. More specifically, the immediate updates made by each participant to its local model can be chained together on the distributed ledger offered by a blockchain such that those model updates are audited. In addition, every model update, be it either local weights or gradients, can be traced to and associated with an individual participant, which helps the detection of tamper attempts and malicious model substitutes. Furthermore, model updates can be chained in a cryptographical way such that their integrity and confidentiality can be guaranteed.
Federated learning has also become an active research topic at the intersection of ML and wireless networks, see, e.g., Habachi et al. [2019], Zhou et al. [2019], Zhu et al. [2018], Samarakoon et al. [2018], Jeong et al. [2018], and particularly for the 5G mobile network and even beyond [Niknam et al., 2019, Letaief et al., 2019, Bennis, 2019, Park et al., 2019]. For instance, the data in wireless networks is usually located at the users and at the network edge, which makes the traditional ML that is based on centralized data collection inapplicable. Federated learning comes as a solution, for addressing not only the data privacy concerns, but also the communication bandwidth, reliability and latency challenges [Bonawitz and Eichner et al., 2019]. Federated learning can also help with building a better wireless network. For example, Niknam et al. [2019] provided an overview of how we can leverage federated learning to address the key challenges and to improve the performance of 5G mobile networks.