Feature-Aware Anisotropic Local Differential Privacy for Utility-Preserving Graph Representation Learning in Metal Additive Manufacturing
Islam, M. S., Bappy, M. M., Tushar, S. R., Arifuzzaman, M.
In-process melt-pool data improve porosity detection, but sharing it across organizations exposes process know-how. We apply local differential privacy at the embedding level and allocate noise anisotropically, guided by Grad-CAM channel importance, before building a layer-wise graph for a GATv2 classifier. The approach retains much of the non-private utility while providing an explicit per-sample privacy guarantee.
Porosity reduces fatigue life and structural integrity in laser-based additive manufacturing. Data-driven detectors benefit from pooled in-process data, yet thermal and optical frames can reveal proprietary process information. In addition, many models classify melt-pool observations independently and ignore their spatial context within a layer.
Privatize each observation locally, before it is shared, and make the noise feature-aware and anisotropic: the privacy budget is concentrated on informative directions instead of being spread uniformly. A graph attention network then exploits neighborhood structure among observations in the same layer.
A CNN encodes each melt-pool frame into a feature vector.
Grad-CAM estimates how strongly each embedding channel contributes to porosity prediction.
Gaussian noise is allocated per coordinate according to channel importance, giving a per-sample privacy guarantee.
Observations within a build layer are linked by k-nearest neighbors in a hybrid space of coordinates and features.
A node-level GATv2 classifies each observation as porous or sound.
Compared against a non-private baseline and a uniform local DP baseline.
The method supports privacy-preserving quality assurance in multi-stakeholder settings, such as AM-as-a-service platforms, where participants benefit from shared models but cannot share raw process data.
Islam, M. S., Bappy, M. M., Tushar, S. R., Arifuzzaman, M.. "Feature-Aware Anisotropic Local Differential Privacy for Utility-Preserving Graph Representation Learning in Metal Additive Manufacturing." Major revision at ASME Journal of Computing and Information Science in Engineering (JCISE). arXiv:2604.05077, 2026.
