Feature-Aware Anisotropic Local Differential Privacy for Utility-Preserving Graph Representation Learning in Metal Additive Manufacturing
Metal additive manufacturing (AM) has revolutionized the fabrication of safety-critical components, primarily due to its ability to produce complex geometries with high precision. However, the reliability of quality assurance in AM processes heavily depends on high-fidelity sensor streams that contain proprietary process information. This reliance on sensitive data poses challenges for collaborative data sharing among manufacturers and researchers.
Existing defect detection models in metal AM often treat melt-pool observations as independent samples. This approach neglects the layer-wise physical couplings that are critical for understanding the manufacturing process. Additionally, conventional privacy-preserving techniques, particularly Local Differential Privacy (LDP), have been shown to lead to significant utility degradation. This degradation arises from the injection of uniform noise across all feature dimensions, which adversely affects the model’s performance.
Introducing FI-LDP-HGAT Framework
To address these interrelated challenges, researchers have proposed a novel computational framework known as FI-LDP-HGAT. This innovative approach integrates two pivotal methodological components:
- Stratified Hierarchical Graph Attention Network (HGAT): This component effectively captures spatial and thermal dependencies across scan tracks and deposited layers, providing a more holistic view of the additive manufacturing process.
- Feature-Importance-Aware Anisotropic Gaussian Mechanism (FI-LDP): This mechanism enhances non-interactive feature privatization by redistributing the privacy budget across embedding coordinates, based on an encoder-derived importance prior.
Unlike traditional isotropic LDP, the FI-LDP mechanism assigns lower noise levels to task-critical thermal signatures while allocating higher noise to redundant dimensions. This allocation maintains formal LDP guarantees without compromising the utility of the data.
Performance and Results
Experiments conducted on a Directed Energy Deposition (DED) porosity dataset have illustrated the effectiveness of the FI-LDP-HGAT framework. The results indicate that:
- FI-LDP-HGAT achieves an impressive 81.5% utility recovery at a moderate privacy budget (epsilon = 4).
- The framework maintains a defect recall rate of 0.762 under stricter privacy conditions (epsilon = 2).
- FI-LDP-HGAT outperforms traditional machine learning models, standard Graph Neural Networks (GNNs), and alternative privacy mechanisms, including Differentially Private Stochastic Gradient Descent (DP-SGD), across all evaluated metrics.
A mechanistic analysis further supports the findings, revealing a strong negative correlation (Spearman = -0.81) between feature importance and noise magnitude. This correlation offers interpretable evidence that the privacy-utility gains achieved by FI-LDP-HGAT are driven by a principled anisotropic allocation strategy.
Conclusion
The FI-LDP-HGAT framework represents a significant advancement in the field of metal additive manufacturing, particularly in the context of defect detection and data privacy. By effectively balancing privacy preservation and utility recovery, this innovative approach has the potential to facilitate collaborative research and improve quality assurance in safety-critical applications.