Feature-Aware Anisotropic Local Differential Privacy for Utility-Preserving Graph Representation Learning in Metal Additive Manufacturing

Md Shafikul Islam, Mahathir Mohammad Bappy, Saifur Rahman Tushar, Md. ArifuzzamanarXiv 2026 · DOI · arXiv

TL;DR

This paper introduces FI-LDP-HGAT, a novel framework that enables privacy-preserving analysis of metal additive manufacturing data. It achieves this by intelligently allocating privacy noise based on feature importance, preserving critical information for defect detection while protecting proprietary process data. This is crucial for enabling collaborative analytics in manufacturing without compromising intellectual property.

What Problem Does This Paper Address?

Metal additive manufacturing (AM) generates vast amounts of sensor data that are vital for quality assurance, particularly for detecting defects like porosity. However, this data often contains proprietary "process fingerprints" that manufacturers are hesitant to share, hindering collaborative efforts and advanced analytics. Existing privacy techniques, like Local Differential Privacy (LDP), often degrade the utility of this data by applying uniform noise across all features, even those critical for defect detection. Furthermore, traditional defect detection methods often treat data points independently, ignoring the crucial physical couplings (like heat accumulation) that influence defect formation in AM. This paper tackles the challenge of preserving both data utility and privacy when performing relational learning on AM process data, specifically for porosity prediction.

Methods

The core of the proposed solution, FI-LDP-HGAT, combines two key components:

  • Feature-Importance-guided Local Differential Privacy (FI-LDP): This is a novel privacy mechanism that moves beyond uniform noise injection. FI-LDP uses a "warmup" phase to estimate the importance of different feature dimensions. It then intelligently redistributes the privacy budget, applying less noise to critical features (like thermal signatures) and more noise to less important ones, while still guaranteeing formal $(\epsilon, \delta)$-LDP. This anisotropic noise allocation is designed to preserve task-relevant information.
  • Stratified Hierarchical Graph Attention Network (HGAT): This is a graph-based learning architecture designed to capture the physical structure of the AM process. It constructs a graph that respects layer-wise dependencies and uses attention mechanisms to learn how to aggregate information from neighboring data points. This approach explicitly models the relational dependencies that influence defect formation, which are missed by traditional methods.

The framework first extracts multimodal features from sensor data, then uses FI-LDP to privatize these features, and finally feeds the privatized features into the HGAT for defect prediction.

Datasets

The authors evaluate their framework on an experimental dataset of Ti-6Al-4V thin-walled structures fabricated via Directed Energy Deposition (DED). This dataset includes high-resolution thermal images and process-state descriptors. A key characteristic of this dataset is its severe class imbalance, with only 4.47% of observations labeled as porous. The data was registered with post-fabrication X-ray Computed Tomography (XCT) pore annotations to establish ground truth.

Key Insights and Ideas

  • Privacy shouldn't be a blunt instrument: The paper demonstrates that applying uniform noise for privacy is inefficient and detrimental to utility, especially when certain features are far more important than others.
  • Feature importance is key to utility-preserving privacy: By understanding which features are most predictive, we can strategically allocate privacy budgets to minimize utility loss.
  • Graph learning is essential for AM defect prediction: Ignoring the relational and physical dependencies in AM data leads to suboptimal performance. HGAT effectively captures these relationships.
  • The privacy-utility trade-off is not fixed: The choice of privacy mechanism and its alignment with data structure significantly impacts how much utility is lost. FI-LDP shows that a more intelligent approach can yield better results.

Key Results

  • Strong Utility Recovery: FI-LDP-HGAT achieves 81.5% utility recovery (measured by F1* score) at a moderate privacy budget ($\epsilon = 4$) compared to the non-private oracle.
  • High Recall Under Strict Privacy: Even under a strict privacy budget ($\epsilon = 2$), the method maintains a defect recall of 0.762, which is crucial for rare-event detection.
  • Outperforms Baselines: FI-LDP-HGAT significantly outperforms standard LDP with HGAT, DP-SGD-style training, and other privacy mechanisms across various metrics.
  • Mechanism Validation: Analysis shows a strong negative correlation (Spearman $\rho = -0.81$) between feature importance and allocated noise, confirming that FI-LDP indeed perturbs informative coordinates less aggressively.
  • Non-Private HGAT Superiority: The non-private HGAT architecture itself achieves a high F1* score of 0.941, demonstrating the effectiveness of the graph-based approach for this task.

Limitations and Future Work

  • Federated Learning Extension: The authors suggest extending FI-LDP-HGAT to multi-facility federated learning to enable joint training of global quality assurance models without sharing proprietary sensor signatures.
  • Physics-Informed Graph Biases: Future work could incorporate more explicit physics priors into graph construction or message passing to improve robustness against privacy noise.
  • Autonomous Control Integration: The framework could be applied to autonomous closed-loop control systems, where privatized representations serve as state variables for real-time defect mitigation.
  • Data Imbalance Mitigation: While porous-targeted augmentation was used, addressing severe class imbalance remains a critical aspect for robust rare-event detection.

The paper acknowledges that while FI-LDP provides a significant improvement over uniform LDP, the utility still degrades with increasing privacy constraints. However, the proposed method offers a more principled and effective way to manage this trade-off.

Summary updated

BibTeX
@misc{islam2026feature,
  author = {Md Shafikul Islam and Mahathir Mohammad Bappy and Saifur Rahman Tushar and Md. Arifuzzaman},
  title = {{Feature-Aware Anisotropic Local Differential Privacy for Utility-Preserving Graph Representation Learning in Metal Additive Manufacturing}},
  year = {2026},
  eprint = {2604.05077},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2604.05077},
}
Figure