Abstract
Federated Learning enables a population of clients, working with a trusted server, to collaboratively learn a shared machine learning model while keeping each client's data within its own local systems. This reduces the risk of exposing sensitive data, but it is still possible to reverse engineer information about a client's private data set from communicated model parameters. Most federated learning systems therefore use differential privacy to introduce noise to the parameters. This adds uncertainty to any attempt to reveal private client data, but also reduces the accuracy of the shared model, limiting the useful scale of privacy-preserving noise. A system can further reduce the coordinating server's ability to recover private client information, without additional accuracy loss, by also including secure multiparty computation. An approach combining both techniques is especially relevant to financial firms as it allows new possibilities for collaborative learning without exposing sensitive client data. This could produce more accurate models for important tasks like optimal trade execution, credit origination, or fraud detection. The key contributions of this paper are: We present a privacy-preserving federated learning protocol to a non-specialist audience, demonstrate it using logistic regression on a real-world credit card fraud data set, and evaluate it using an open-source simulation platform which we have adapted for the development of federated learning systems.
Introduction
Financial institutions possess vast amounts of sensitive data that could significantly improve machine learning models for fraud detection or credit scoring if pooled. However, regulatory requirements and privacy concerns often prohibit direct data sharing. While Federated Learning (FL) allows for local model training on decentralized data, it is not inherently private. An "honest-but-curious" central server can potentially reconstruct private local data from the model updates (weights) sent by clients.
To mitigate this, Differential Privacy (DP) is often employed by adding noise to updates. However, DP introduces a trade-off: more noise increases privacy but degrades model utility. Secure Multi-Party Computation (MPC) offers another layer of protection by hiding individual updates from the server, yet it remains vulnerable to client collusion. This paper proposes a "defense-in-depth" framework that integrates FL, DP, and MPC to provide robust privacy guarantees without sacrificing the accuracy required for high-stakes financial applications.
Methodology: The $\Pi_{PPFL}$ Protocol
The proposed protocol, $\Pi_{PPFL}$, layers three distinct technologies to ensure data privacy:
- Federated Logistic Regression: Clients compute local model weights ($w$) based on their private datasets.
- Differential Privacy (DP): Before sharing, clients add Laplacian noise to their weights. This ensures that even if the underlying weights are exposed (e.g., via collusion), the individual data points remain protected.
- Secure Multi-Party Computation (MPC): Based on the Bonawitz protocol, clients exchange secret random offsets. As illustrated in Figure 1, clients pair up; one adds an offset while the other subtracts it. When the server aggregates these masked weights ($\bar{w}$), the offsets cancel out, allowing the server to compute the true average without ever seeing an individual's unmasked (but still noisy) weights.
To evaluate the protocol realistically, the authors utilized ABIDES, an agent-based discrete event simulator. Unlike standard simulations, ABIDES accounts for network latency, message queuing, and jitter, providing a high-fidelity assessment of the protocol's operational performance.
Experimental Setup
The protocol was tested using the Kaggle Credit Card Fraud (2013) dataset, which contains 284,807 transactions. A critical challenge of this dataset is its extreme class imbalance: only 0.17% of transactions are fraudulent. In such scenarios, standard accuracy is a poor metric, as a model predicting "no fraud" for every case would achieve 99.8% accuracy while failing its primary purpose. Consequently, the authors used the Matthews Correlation Coefficient (MCC) to measure performance, as it remains informative even for highly imbalanced classes.
Results and Discussion
The experiments revealed several key insights regarding the scalability and efficacy of the privacy-preserving framework.
Performance and Privacy Trade-offs
A primary finding is that the "privacy budget" improves with the number of participants. As the client population grows, the amount of noise required per individual to maintain a global privacy level decreases. Figure 3 demonstrates that for populations exceeding 200 clients, the model achieves high MCC scores even with a small privacy loss parameter ($\epsilon$). This suggests that large-scale federations can achieve performance comparable to non-private centralized training.
Scalability and Efficiency
The simulation results indicated that the protocol is computationally feasible for real-world deployment. A federation of 500 parties could complete a training iteration in approximately 17 seconds. To further optimize communication, the authors demonstrated that clients do not need to communicate with every other participant. By limiting key exchanges to a "neighborhood" of size $\log(N)$, the protocol maintains security while significantly reducing the $O(N^2)$ communication bottleneck typical of standard MPC.
Security and Collusion Resistance
The combination of DP and MPC provides a robust defense. Even in a worst-case scenario where $n-1$ clients collude to isolate a single honest participant's contribution, they would only recover a version of the weights protected by Laplacian noise. This multi-layered approach ensures that no single point of failure—be it the server or a group of malicious clients—can compromise individual privacy.
Conclusion and Future Work
This paper demonstrates a practical, high-performance framework for privacy-preserving federated learning in the financial sector. By combining DP and MPC, the $\Pi_{PPFL}$ protocol enables collaborative model training on sensitive data with minimal accuracy loss.
Future research will focus on extending this framework to more complex models, such as deep neural networks, and addressing "active" adversaries who might attempt to poison the global model with malicious updates. Additionally, further optimizations to the initial Diffie-Hellman setup phase could enhance scalability for even larger client populations.
Summary updated