Explainable Differential Privacy-Hyperdimensional Computing for Balancing Privacy and Transparency in Additive Manufacturing Monitoring

Fardin Jalil Piran, Prathyush Poduval, Hamza Errahmouni Barkam, Mohsen Imani, Farhad Imani

Introduction

Additive Manufacturing (AM), or industrial 3D printing, is rapidly transitioning toward real-time "in-situ" monitoring to identify defects such as porosity or incomplete fusion during the build process. While Machine Learning (ML) models—particularly deep learning architectures—excel at identifying these defects from high-speed camera feeds, they introduce significant security risks. Training data often contains proprietary part designs, material compositions, and process parameters that can be reconstructed by attackers through "model inversion" attacks.

To mitigate these risks, researchers employ Differential Privacy (DP), a technique that injects mathematical noise into the training process. However, in traditional "black-box" AI models, it is notoriously difficult to predict how much noise can be added before classification accuracy collapses. This paper introduces DP-HD, a framework that leverages Hyperdimensional Computing (HD) and a novel Signal-to-Noise Ratio (SNR) metric to provide a predictable, transparent, and high-performance solution for secure AM monitoring.

The DP-HD Framework

The DP-HD framework deviates from standard neural networks by utilizing Hyperdimensional Computing, a brain-inspired paradigm that represents data as massive vectors (typically 10,000+ dimensions). The framework's architecture is built on four pillars:

  1. Holographic Representation: Unlike traditional neural networks where specific weights may hold critical information, HD distributes information across the entire vector. This redundancy makes the model inherently more robust to noise.
  2. Direct DP Integration: Gaussian noise is injected directly into the "class hypervectors" during training. This approach simplifies the privacy-preserving mechanism compared to the complex gradient-clipping required in deep learning.
  3. Explainable SNR Metric: The authors developed a Signal-to-Noise Ratio (SNR) metric to quantify the contribution of training data relative to the DP noise. This allows engineers to mathematically forecast model accuracy and determine the optimal privacy budget ($\epsilon$) before training begins, eliminating the typical trial-and-error process.
  4. Inference Privacy: To protect data during the prediction phase, the framework identifies and drops "low-variance" dimensions in query vectors. This prevents attackers from reconstructing original images from the model's queries without impacting classification performance.

Experimental Methodology

The framework was validated using both real-world and synthetic datasets:

  • NIST AM Data: Thermal signature data from an EOS M270 laser powder bed fusion (LPBF) machine. The dataset includes high-speed camera footage of 45 build layers, focusing on overhangs and bulk features.
  • Synthetic Sensitivity Analysis: Used to test how varying class distances and sample spreads affect the optimal encoding strategy, specifically the standard deviation ($\sigma_b$) used for random basis vectors.

Key Results and Insights

The DP-HD framework demonstrated superior performance across three critical metrics: accuracy under privacy constraints, computational efficiency, and data scarcity robustness.

Privacy vs. Accuracy

Under strict privacy constraints ($\epsilon=1$), DP-HD achieved a remarkable 94.43% accuracy. In contrast, state-of-the-art deep learning models like ResNet50 and DenseNet201 saw their accuracy plummet to between 30% and 52% under the same conditions. This highlights HD's unique ability to maintain a strong signal even when heavily masked by noise.

Inference Speed

In-situ monitoring requires millisecond-latency to allow for real-time process corrections. DP-HD processed 1,610 samples in just 58.15ms. This performance is approximately 13x faster than ResNet50 (772ms) and 34x faster than DenseNet201 (2027ms), making it the only viable candidate among the tested models for high-speed industrial applications.

Robustness to Small Datasets

Differential Privacy typically requires very large datasets to overcome the "noise floor." However, DP-HD maintained high accuracy even with small training sets (e.g., 1,000 samples), significantly outperforming DP-SGD (Stochastic Gradient Descent) methods used in traditional deep learning.

Conclusion and Future Work

DP-HD transforms privacy-preserving machine learning from a "guessing game" into an engineering calculation. By utilizing the inherent robustness of Hyperdimensional Computing and the transparency of the SNR metric, the framework allows manufacturers to protect their intellectual property without sacrificing the accuracy of defect detection.

Future research will focus on automating the selection of encoding parameters (such as the optimal $\sigma_b$) based on raw data distributions and validating the framework in broader industrial collaborations, including aerospace and automotive applications. Additionally, the authors plan to explore the integration of DP-HD into Federated Learning structures to enable secure, cross-facility model training.

Summary updated

Abstract

Machine Learning (ML) models integrated with in-situ sensing offer transformative solutions for defect detection in Additive Manufacturing (AM), but this integration brings critical challenges in safeguarding sensitive data, such as part designs and material compositions. Differential Privacy (DP), which introduces mathematically controlled noise, provides a balance between data utility and privacy. However, black-box Artificial Intelligence (AI) models often obscure how this noise impacts model accuracy, complicating the optimization of privacy-accuracy trade-offs. This study introduces the Differential Privacy-Hyperdimensional Computing (DP-HD) framework, a novel approach combining Explainable AI (XAI) and vector symbolic paradigms to quantify and predict noise effects on accuracy using a Signal-to-Noise Ratio (SNR) metric. DP-HD enables precise tuning of DP noise levels, ensuring an optimal balance between privacy and performance. The framework has been validated using real-world AM data, demonstrating its applicability to industrial environments. Experimental results demonstrate DP-HD's capability to achieve state-of-the-art accuracy (94.43%) with robust privacy protections in anomaly detection for AM, even under significant noise conditions. Beyond AM, DP-HD holds substantial promise for broader applications in privacy-sensitive domains such as healthcare, financial services, and government data management, where securing sensitive data while maintaining high ML performance is paramount.
BibTeX
@article{piran2025explainable,
  author = {Fardin Jalil Piran and Prathyush Poduval and Hamza Errahmouni Barkam and Mohsen Imani and Farhad Imani},
  title = {{Explainable Differential Privacy-Hyperdimensional Computing for Balancing Privacy and Transparency in Additive Manufacturing Monitoring}},
  journal = {Engineering Applications of AI},
  year = {2025},
  doi = {10.1016/j.engappai.2025.110282},
}
Figure