Abstract
Deep Neural Networks (DNNs) are powerful tools for applications like fraud detection, but "data isolation"—the inability of parties to share sensitive data—remains a significant barrier. While cryptographic techniques offer privacy, they often fail to scale to industrial datasets. This paper introduces $P^2N^2$, a privacy-preserving neural network learning paradigm that balances security and scalability. By splitting the computation graph and employing a novel "defender" mechanism, $P^2N^2$ enables collaborative training on millions of records with minimal accuracy loss and high efficiency.
Introduction: The Data Isolation Challenge
In sectors like finance and insurance, data is often siloed. While combining datasets from multiple institutions could significantly improve model performance, strict privacy regulations like GDPR prevent direct sharing. Traditional cryptographic solutions, such as Homomorphic Encryption (HE) or Secure Multi-Party Computation (MPC), provide strong security but introduce massive computational and communication overheads, making them impractical for deep learning on industrial-scale datasets.
The $P^2N^2$ Framework
The proposed Privacy Preserving Neural Network ($P^2N^2$) architecture addresses the scalability-privacy trade-off through a hybrid approach.
Computation Splitting and Secret Sharing
The core idea is to divide the DNN's computation graph into private and public components. Data holders locally process the input layers and the final label/prediction layers. The computationally intensive middle hidden layers are offloaded to a neutral third-party server. To protect raw features during the initial combination, the first hidden layer is computed via secret sharing, ensuring that no party reveals their raw data to others.
The Defender Mechanism
A critical risk in split learning is "inversion attacks," where a curious server might attempt to reconstruct raw inputs from the hidden layer activations it receives. $P^2N^2$ introduces an adversarial defender mechanism. During training, the model is optimized not only for the primary task (e.g., fraud detection) but also to minimize the ability of an internal "attacker" to reconstruct the original input. This adversarial loss forces the network to learn representations that are task-relevant but privacy-preserving.
Experimental Evaluation
The framework was tested on real-world datasets, including Kaggle's Fraud Detection and Financial Distress datasets, as well as MNIST for visual verification.
Key Results
- Accuracy: On the Fraud Detection task, $P^2N^2$ achieved an AUC of 0.9231, nearly matching the 0.9270 achieved by a standard non-private model.
- Scalability: While the secret sharing step makes $P^2N^2$ slower than plaintext training, it remains orders of magnitude faster than pure cryptographic methods, proving its viability for industrial use.
- Training Stability: The model demonstrates smooth convergence, as shown in the training and test loss curves.
Conclusion
$P^2N^2$ represents a practical step forward for privacy-preserving machine learning in industry. By strategically applying cryptographic techniques and leveraging adversarial training for privacy, it overcomes the scalability hurdles of traditional methods. Future work will focus on hardening the system against malicious (rather than just semi-honest) actors and deploying the framework in live production environments at Ant Financial.
Summary updated