Feature-Aware Anisotropic Local Differential Privacy for Utility-Preserving Graph Representation Learning in Metal Additive Manufacturing
arXiv:2604.05077v1 Announce Type: new Abstract: Metal additive manufacturing (AM) enables the fabrication of safety-critical components, but reliable quality assurance depends on high-fidelity sensor streams containing proprietary process information, limiting collaborative data sharing. Existing defect-detection models typically treat melt-pool observations as independent samples, ignoring layer-wise physical couplings. Moreover, conventional privacy-preserving techniques, particularly Local Differential Privacy (LDP), lead to severe utility degradation because they inject uniform noise across all feature dimensions. To address these interrelated challenges, we propose FI-LDP-HGAT. This computational framework combines two methodological components: a stratified Hierarchical Graph Attention Network (HGAT) that captures spatial and thermal dependencies across scan tracks and deposited layers, and a feature-importance-aware anisotropic Gaussian mechanism (FI-LDP) for non-interactive fe
arXiv:2604.05077v1 Announce Type: new Abstract: Metal additive manufacturing (AM) enables the fabrication of safety-critical components, but reliable quality assurance depends on high-fidelity sensor streams containing proprietary process information, limiting collaborative data sharing. Existing defect-detection models typically treat melt-pool observations as independent samples, ignoring layer-wise physical couplings. Moreover, conventional privacy-preserving techniques, particularly Local Differential Privacy (LDP), lead to severe utility degradation because they inject uniform noise across all feature dimensions. To address these interrelated challenges, we propose FI-LDP-HGAT. This computational framework combines two methodological components: a stratified Hierarchical Graph Attention Network (HGAT) that captures spatial and thermal dependencies across scan tracks and deposited layers, and a feature-importance-aware anisotropic Gaussian mechanism (FI-LDP) for non-interactive feature privatization. Unlike isotropic LDP, FI-LDP redistributes the privacy budget across embedding coordinates using an encoder-derived importance prior, assigning lower noise to task-critical thermal signatures and higher noise to redundant dimensions while maintaining formal LDP guarantees. Experiments on a Directed Energy Deposition (DED) porosity dataset demonstrate that FI-LDP-HGAT achieves 81.5% utility recovery at a moderate privacy budget (epsilon = 4) and maintains defect recall of 0.762 under strict privacy (epsilon = 2), while outperforming classical ML, standard GNNs, and alternative privacy mechanisms, including DP-SGD across all evaluated metrics. Mechanistic analysis confirms a strong negative correlation (Spearman = -0.81) between feature importance and noise magnitude, providing interpretable evidence that the privacy-utility gains are driven by principled anisotropic allocation.
Executive Summary
The article presents FI-LDP-HGAT, a novel framework addressing critical challenges in metal additive manufacturing (AM) by integrating a feature-importance-aware anisotropic Gaussian Local Differential Privacy (LDP) mechanism with a Hierarchical Graph Attention Network (HGAT). The framework overcomes limitations of conventional LDP, which degrades utility by uniformly injecting noise, and traditional defect-detection models that ignore layer-wise physical couplings in melt-pool observations. By redistributing the privacy budget based on feature importance, FI-LDP-HGAT achieves substantial utility recovery (81.5% at epsilon=4) and maintains defect recall (0.762 at epsilon=2) while outperforming classical ML, GNNs, and DP-SGD. The study demonstrates a strong negative correlation between feature importance and noise magnitude, highlighting the principled trade-off between privacy and utility. The work advances both data privacy and graph representation learning in industrial AM contexts, offering a scalable solution for collaborative yet secure data sharing in safety-critical applications.
Key Points
- ▸ FI-LDP-HGAT introduces an anisotropic Gaussian mechanism for LDP that prioritizes task-critical thermal features while suppressing noise on redundant dimensions, preserving utility without compromising privacy guarantees.
- ▸ The stratified HGAT architecture captures spatial and thermal dependencies across scan tracks and layers in directed energy deposition (DED), addressing limitations of prior models that treat melt-pool observations as independent samples.
- ▸ Empirical evaluation on a DED porosity dataset demonstrates superior performance—81.5% utility recovery at epsilon=4 and defect recall of 0.762 at epsilon=2—compared to classical ML, standard GNNs, and alternative privacy mechanisms including DP-SGD.
- ▸ Mechanistic analysis reveals a strong negative Spearman correlation (-0.81) between feature importance and noise magnitude, validating the effectiveness of the anisotropic allocation strategy.
- ▸ The framework enables secure, collaborative data sharing in proprietary manufacturing environments, addressing a key barrier to quality assurance in safety-critical AM components.
Merits
Innovative Integration of Privacy and Utility
The fusion of a feature-aware anisotropic LDP mechanism with a hierarchical graph attention network represents a significant methodological advance, enabling formal privacy guarantees without sacrificing task performance.
Empirical Rigor and Reproducibility
The study provides strong empirical validation through a large-scale DED dataset, with detailed benchmarking against multiple baselines and ablation studies that isolate the contributions of each component.
Industrial Relevance and Impact
The work directly addresses a critical need in metal AM—secure data sharing for collaborative quality assurance—while maintaining high fidelity in defect detection, making it highly relevant for real-world deployment.
Interpretability
The observed correlation between feature importance and noise allocation provides clear, mechanistic insight into how privacy-utility trade-offs are managed, enhancing transparency and trust in the system.
Demerits
Assumption of Feature Importance Prior
The framework relies on an encoder-derived importance prior to guide noise allocation. While effective, this assumes the encoder’s learned importance is accurate and task-aligned, which may not hold under adversarial conditions or distribution shift.
Limited Generalizability Claims
The study is validated on a single DED dataset. While results are promising, broader applicability to other AM processes (e.g., powder bed fusion) or sensor modalities (e.g., acoustic, optical) remains unverified.
Privacy Budget Sensitivity
While moderate privacy budgets (epsilon=2–4) are evaluated, the behavior at extreme privacy levels (e.g., epsilon <1) or in high-dimensional settings is not explored, leaving open questions about scalability in ultra-private regimes.
Computational Overhead
The HGAT and anisotropic LDP mechanisms introduce additional computational complexity compared to standard LDP or simple GNNs. The trade-off between performance gains and resource costs is not quantified in detail.
Expert Commentary
This article represents a sophisticated convergence of machine learning, privacy engineering, and industrial informatics, offering a compelling solution to a pressing challenge in metal additive manufacturing. The authors correctly identify a fundamental tension between privacy and utility in collaborative quality assurance, where proprietary sensor data must be shared for training robust defect-detection models. By leveraging an anisotropic LDP mechanism guided by feature importance, they transcend the limitations of isotropic noise injection, achieving a rare balance between formal privacy guarantees and high task performance. The integration of HGAT further addresses a critical gap in existing models by capturing the hierarchical dependencies inherent in AM processes. While the reliance on a learned importance prior introduces a degree of fragility, the strong empirical results and interpretability of the noise allocation mechanism significantly mitigate this concern. This work not only advances the technical frontier of privacy-preserving graph learning but also provides a blueprint for secure, collaborative AI in industrial settings where trust and performance are non-negotiable. Future work should explore robustness under distribution shift and extend the framework to multimodal sensor fusion, but the current contribution stands as a landmark in industrial AI privacy.
Recommendations
- ✓ Future research should validate FI-LDP-HGAT on diverse AM processes and sensor modalities to establish broader generalizability, particularly in powder bed fusion and multi-physics monitoring systems.
- ✓ Investigate the robustness of the feature-importance prior under adversarial conditions or distribution shift, potentially incorporating techniques from adversarially robust learning to harden the privacy mechanism.
- ✓ Develop standardized evaluation protocols for privacy-preserving quality assurance systems in manufacturing, including benchmarks for utility recovery, defect recall, and computational efficiency across varying epsilon budgets.
- ✓ Explore hybrid privacy-preserving paradigms, such as combining FL with FI-LDP-HGAT, to further enhance scalability and reduce reliance on centralized data aggregation in collaborative manufacturing networks.
- ✓ Engage with industry partners to pilot the framework in real-world AM facilities, assessing operational feasibility, regulatory compliance, and long-term impact on defect reduction and process optimization.
Sources
Original: arXiv - cs.LG