Abstract
Federated Learning (FL) is emerging as the technological backbone for the "Open Banking" ecosystem, enabling financial institutions to collaboratively train powerful AI models while adhering to strict privacy regulations like GDPR. This paper identifies the unique challenges of applying FL to banking—such as data heterogeneity and "pay-per-access" constraints—and surveys state-of-the-art solutions including clustered learning, knowledge distillation, and one-class classification.
Introduction
Open banking aims to give customers ownership of their data, allowing them to share it with third-party providers to access superior financial services. However, centralizing this data for machine learning (ML) is risky and often illegal under frameworks like the EU’s General Data Protection Regulation (GDPR). Federated Learning (FL) provides a decentralized alternative where models are trained locally, and only model updates are shared with a central server.
Challenges in the Banking Sector
Standard FL is not a "plug-and-play" fix for banking due to several domain-specific hurdles:
- Statistical Heterogeneity: Banking data is "non-IID" (not independently and identically distributed); for instance, a student's spending habits differ significantly from a corporation's.
- Model Heterogeneity: Different banks may utilize different neural network architectures.
- Access Limits: In a data marketplace, accessing a user's profile might be charged per use, requiring "few-shot" learning to minimize communication.
- The One-Class Problem: Most users have never committed fraud, meaning local data often contains only "positive" (normal) examples, complicating binary classification for anomaly detection.
Methodological Framework
The paper categorizes mathematical approaches to these challenges, focusing on personalization and efficiency.
Clustered Federated Learning
Instead of one global model, participants are grouped by data similarity. Techniques such as using the cosine similarity of gradient updates allow the system to iteratively split participants into clusters that share a unique global model tailored to their specific data distribution.
Personalized Modeling and Knowledge Distillation
To prevent global models from being too generic, the authors discuss Personalization Layers (sharing representation layers but keeping decision layers local) and Mixture Models. To handle different model architectures, Knowledge Distillation (KD) is used to "distill" knowledge from complex models into simpler ones by exchanging prediction probabilities on public datasets rather than raw model parameters.
Efficiency and One-Class FL
To minimize financial costs associated with data access, the paper discusses ensemble learning techniques that allow a server to learn a global model in as little as one round. For fraud detection, the authors suggest embedding one-class classification algorithms, such as auto-encoders, into the FL framework to handle the lack of negative examples at the local level.
Key Insights and Market Implications
- FL as a Marketplace Enabler: FL is framed as a mechanism for a "data marketplace" where data remains local, but its utility is sold.
- Privacy Risks: The paper warns that exchanging gradients can lead to "gradient leakage," making Differential Privacy and Secure Aggregation mandatory companions to FL in banking.
- The "Pay-per-Access" Reality: Unlike consumer-grade FL, open banking may involve participants charging for every coordinator "ping," making communication efficiency a financial necessity.
Conclusion and Future Work
While FL offers a promising path forward, several challenges remain. Incentive management using game theory is required to encourage data sharing. Furthermore, banking data is a "moving target," necessitating incremental or lifelong learning strategies to capture evolving customer behavior. Finally, the threat of adversarial "poisoning" attacks remains an open area for research, with blockchain-based ledgers proposed as a potential remedy.
Summary updated