Abstract
Researchers at IBM Research – Africa have developed a blockchain-based platform that utilizes Variational Auto-Encoders (VAEs) to generate synthetic financial datasets. This approach addresses the "Privacy-Utility-Fairness" trilemma by enabling institutions to share data for bias auditing and model training while protecting individual privacy. By breaking the one-to-one link between real individuals and data points, the system successfully preserves the "group bias" characteristics of the original data, though it faces challenges in maintaining individual fairness metrics.
Introduction
In the financial sector, machine learning models require massive datasets to ensure accuracy and fairness. However, this data often contains Sensitive Personal Information (SPI) subject to strict privacy regulations like the "Right to be Forgotten." Traditional anonymization techniques, such as masking or encryption, are frequently insufficient as they remain vulnerable to de-identification by malicious actors. Furthermore, small-to-medium enterprises (SMEs) often face data scarcity, hindering the development of robust models. This research tackles the challenge of sharing data for auditing and expansion without compromising individual privacy or losing the statistical nuances—including inherent biases—necessary for study.
System Architecture: The Trusted Model Executor
The authors propose the Trusted Model Executor (TME), a platform integrated with a blockchain backend to manage the entire lifecycle of a machine learning model with transparency and immutability.
The TME ensures that all interactions with models and data are recorded on a ledger, providing a verifiable audit trail. This architecture supports a collaborative environment where developers can upload models and data for analysis, synthesis, and bias detection within a secure framework.
Methodology: Data Synthesis via VAEs
The core technical innovation for data sharing is the use of Variational Auto-Encoders (VAEs). Unlike standard anonymization, a VAE learns the underlying probabilistic distribution of a dataset and maps it to a "latent space." By sampling from this space, the system generates entirely new, synthetic records.
This method severs the one-to-one relationship between real records and synthetic ones, making re-identification nearly impossible. Crucially, for a synthetic dataset to be useful for auditing, it must retain the biases of the original. If the original data discriminated against a specific group, the synthetic data must reflect that same disparity to allow for the development of effective mitigation strategies.
Experimental Evaluation and Results
The study utilized two benchmark tabular datasets: the US Adult Census Dataset (predicting income thresholds) and the Bank of Portugal Dataset (predicting term deposit subscriptions). To test bias detection, these datasets were intentionally skewed through subsampling before generating synthetic versions.
Quantitatively, the correlation between the real and synthetic datasets for group bias metrics was remarkably high. The Statistical Parity Difference showed a 0.96 correlation, while Disparate Impact reached a 0.98 correlation. These results confirm that the VAE approach is highly effective at capturing "group fairness."
However, the Consistency metric—a proxy for individual fairness—showed a weak negative correlation of -0.34. This suggests that while the "big picture" of the data is preserved, the local relationships between similar data points are altered during the synthesis process.
Conclusion and Future Work
The primary limitation of the current VAE-based approach is the loss of individual fairness and consistency. The synthesis process effectively "shuffles" the local neighborhood of data points, which may limit its suitability for granular, individual-level analysis. Future research will focus on refining VAE architectures to better capture these individual nuances and testing the platform across more diverse datasets to improve performance consistency.
Summary updated