Source-linked AI summary
Federated Learning for Open Banking
Guodong Long, Yue Tan, Jing Jiang, Chengqi Zhang
TL;DR
Open banking creates opportunities for data marketplaces and financial services, but sharing sensitive customer data raises privacy and governance challenges. This chapter examines federated learning as decentralized training without direct private-data collection and discusses solutions for heterogeneity, access limits, one-class data, and incentives. It concludes by surveying practical challenges and corresponding techniques for applying federated learning in open banking.
Problem
Applying federated learning to open banking is constrained by privacy risks, heterogeneous users and models, limited data access, one-class data, and incentive management.
Method
The chapter analyzes open-banking federated-learning challenges and surveys corresponding techniques, using a server-and-participant decentralized training framework.
Results
The chapter identifies statistical heterogeneity, model heterogeneity, access limits, and one-class problems as under-discussed open-banking challenges and explores solutions for them.
Takeaways & Limitations
Federated learning combines naturally with open-banking data marketplaces because models can be trained across distributed data ownership without directly collecting private data.
Takeaways & Limitations
The chapter does not discuss privacy-preserving techniques and data encryption in detail because these topics are covered in existing literature.
Abstract
from arXiv · showhide
Open banking enables individual customers to own their banking data, which provides fundamental support for the boosting of a new ecosystem of data marketplaces and financial services. In the near future, it is foreseeable to have decentralized data ownership in the finance sector using federated learning. This is a just-in-time technology that can learn intelligent models in a decentralized training manner. The most attractive aspect of federated learning is its ability to decompose model training into a centralized server and distributed nodes without collecting private data. This kind of decomposed learning framework has great potential to protect users' privacy and sensitive data. Therefore, federated learning combines naturally with an open banking data marketplaces. This chapter will discuss the possible challenges for applying federated learning in the context of open banking, and the corresponding solutions have been explored as well.
1 Introduction
Open banking expands banking into an API-enabled financial-services ecosystem and may support data marketplaces, but sharing sensitive customer data creates privacy and governance risks. Federated learning is presented as a decentralized approach suited to this setting, although practical adaptations are needed.
- Open banking uses APIs, transparency options, and open-source technology to connect third-party services with financial institutions and support data marketplaces.
- Sharing banking data can improve customer experience, revenue streams, and services for underserved markets, but creates privacy, security, and governance concerns.
- Federated learning collaboratively trains AI models across distributed data owners without directly collecting users’ private data.
- Applying federated learning to open banking requires addressing user heterogeneity, limited data access, narrow individual-data scope, and contributor incentives.
2 Applications of open banking
Open banking supports data-driven financial applications and marketplaces, while federated learning offers decentralized training without direct private-data access. The chapter identifies privacy, incentives, heterogeneity, access limits, and one-class learning as practical challenges.
- Open banking applications: Open banking is a collaborative API-based model that enables unaffiliated parties to share banking data for enhanced marketplace capabilities.
- Open banking applications: Applications include fraud detection, credit assessment, customer retention, personalized services, payment information exchange, and integrated financial tools.
- Federated learning challenges: Federated learning avoids direct private-data access, but exchanged parameters or gradients may still leak information and require privacy protection.
- Federated learning challenges: Open banking federated learning must address incentives, heterogeneous data and models, limited profile access, and one-class participant data.
3 Problem formulation
Federated learning separates training between a coordinating server and participant nodes. Participants fine-tune a shared model locally, while the global objective weights and minimizes their losses.
- Federated learning decomposes training between a server coordinator and distributed participants connected through a designed mechanism.
- Each participant fine-tunes globally shared parameters W into local parameters W_i using its own dataset D_i.
- The global model is optimized to minimize the weighted total loss across participants.
4 Statistical heterogeneity in federated learning
Statistical heterogeneity arises because participants have different data distributions, making a single global model inadequate for diverse users. Proposed responses include clustered models and personalized models that combine shared and participant-specific information.
- 4 Statistical heterogeneity in federated learning: Statistical heterogeneity is a non-IID setting where participants generate data from different distributions, challenging vanilla federated optimization.Differences in p(x), p(y), p(x|y), or p(y|x) can create inconsistent learning across participants.
- 4.1 Clustered federated learning: Clustered federated learning increases the number of global models so participants with similar distributions share one model.Clustering can be based on model similarity, gradient-update cosine similarity, or participant-specific loss.
- 4.2 Personalized modelling: Personalized federated learning gives each participant a model combining commonly shared information with personalized information.Approaches include data or model interpolation, personalized layers, mixture models, and constrained personalization.
- 4.1 Clustered federated learning: CNN-based model comparison requires functionality-aware neuron matching because equivalent neurons may occupy different indices across models.Such matching can improve performance and support clustering-based federated learning.
- 4.2 Personalized modelling: Personalization can overfit when participants have few local instances, so transfer learning, domain adaptation, or regularization may be used to control local adaptation.Regularization limits the distance of personalized changes from the global model.
5 Model heterogeneity
Model heterogeneity occurs when participants use different architectures, preventing direct aggregation of their local models. Knowledge distillation is presented as a way to transfer information into a common or smaller model, though one discussed solution assumes pre-training.
- 5 Model heterogeneity: Different participant architectures make the local models and global model incompatible with the standard aggregation operator.Aggregation requires transforming heterogeneous models into homogeneous models.
- 5 Model heterogeneity: Knowledge distillation transfers information from a teacher model to a simpler student model by matching predicted outputs while considering ground truth.The loss combines a soft loss between predicted-label distributions with a ground-truth loss.
- 5 Model heterogeneity: A decentralized distillation framework can exchange prediction probabilities on a shared public dataset while each participant trains with private and public data.The coordinator receives participant predictions rather than requiring direct access to private datasets.
- 5 Model heterogeneity: Federated distillation can exchange model outputs instead of parameters, improving communication efficiency and potentially replacing sensitive parameters.The cited approaches include semi-supervised distillation and on-device machine learning.
- 5 Model heterogeneity: A personalization-oriented distillation solution assumes a pre-trained federated model, limiting its role to deployment or personalization rather than federated learning.The global model acts as teacher and the adapted client model as student.
6 Limited number of uses of participants
Open banking may restrict how often a participant’s data can be accessed, while customer profiles change over time. Few-shot federated learning addresses this by minimizing communication rounds and improving data-use efficiency.
- 6 Limited number of uses of participants: Pay-per-access data marketplaces charge participants according to the number of times their data are used during federated training.A process requiring three local training sessions would charge the participant three times.
- 6 Limited number of uses of participants: Because banking profiles change dynamically, federated learning may require incremental or lifelong learning strategies.These strategies aim to capture evolving customer banking activities.
- 6 Limited number of uses of participants: Few-shot federated learning models the setting as very limited communication between the server and each participant.The coordinator trains local models using global parameters at a particular moment and seeks highly efficient communication.
- 6 Limited number of uses of participants: One-shot federated learning enables a central server to learn a global model in a single communication round using ensemble learning and knowledge aggregation.The approach is described as leveraging information efficiently across networks with thousands of devices.
7 Only positive labels in each participant
In open banking fraud detection, each participant may observe only one label class even though the overall task is binary. Solutions combine one-class algorithms with system-level weighting and incentives, while blockchain-based methods address poisoning risks.
- 7 Only positive labels in each participant: A participant’s financial data may contain only fraud or not-fraud labels, with most users belonging to the not-fraud class.This one-class setting can make individually trained personalized models inaccurate.
- 7 Only positive labels in each participant: The overall federated task is binary although each participant may face only a one-class learning task.The mismatch between local and overall label structure creates the one-class challenge.
- 7 Only positive labels in each participant: One-class solutions operate at the task level through specialized classification algorithms or at the system level through adjusted weights and incentive mechanisms.One-class classification is also called anomaly detection in fraud detection applications.
- 7 Only positive labels in each participant: A blockchain-based federated learning method chains auto-encoder updates on a distributed ledger to address local-model poisoning and delayed attack detection.The trained auto-encoder is used to recognize test samples.
8 Summary
The chapter examines practical challenges that arise when applying federated learning to open banking and surveys corresponding solutions. It emphasizes four underexplored areas: statistical heterogeneity, model heterogeneity, access limits, and one-class problems.
- The chapter focuses on statistical heterogeneity, model heterogeneity, access limits, and one-class problems in open-banking federated learning.These challenges are described as rarely discussed elsewhere.
- It explores various solutions to address the practical challenges of applying federated learning in open banking.
- The chapter considers how federated learning may advance in real-world open-banking scenarios.