[Paper Review] RES-HD: Resilient Intelligent Fault Diagnosis Against Adversarial Attacks Using Hyper-Dimensional Computing

Onat Güngör, Tajana Rosing+1 more|arXiv (Cornell University)|Mar 14, 2022

Advanced Memory and Neural Computing4 citations

TL;DR

This paper proposes RES-HD, a hyper-dimensional computing (HDC)-based fault diagnosis system that achieves up to 67.5% higher resilience against black-box adversarial attacks compared to state-of-the-art deep learning models, while training 25.1× faster. It uses a transferable attack framework with a substitute WDCNN model to generate adversarial examples, demonstrating HDC's robustness and efficiency in industrial IoT environments.

ABSTRACT

Industrial Internet of Things (I-IoT) enables fully automated production systems by continuously monitoring devices and analyzing collected data. Machine learning methods are commonly utilized for data analytics in such systems. Cyber-attacks are a grave threat to I-IoT as they can manipulate legitimate inputs, corrupting ML predictions and causing disruptions in the production systems. Hyper-dimensional computing (HDC) is a brain-inspired machine learning method that has been shown to be sufficiently accurate while being extremely robust, fast, and energy-efficient. In this work, we use HDC for intelligent fault diagnosis against different adversarial attacks. Our black-box adversarial attacks first train a substitute model and create perturbed test instances using this trained model. These examples are then transferred to the target models. The change in the classification accuracy is measured as the difference before and after the attacks. This change measures the resiliency of a learning method. Our experiments show that HDC leads to a more resilient and lightweight learning solution than the state-of-the-art deep learning methods. HDC has up to 67.5% higher resiliency compared to the state-of-the-art methods while being up to 25.1% faster to train.

Motivation & Objective

• Address the growing threat of adversarial attacks on machine learning models in Industrial IoT (I-IoT) systems, which can disrupt predictive maintenance and fault diagnosis.
• Overcome the vulnerability of deep learning models to input perturbations that degrade classification accuracy under adversarial conditions.
• Develop a lightweight, resilient, and efficient learning solution for intelligent fault diagnosis (IFD) in resource-constrained I-IoT environments.
• Demonstrate that hyper-dimensional computing (HDC) offers superior robustness and computational efficiency compared to state-of-the-art deep learning methods under black-box adversarial attacks.

Proposed method

• Employ a black-box adversarial attack strategy based on transferability: train a substitute wide deep convolutional neural network (WDCNN) to generate adversarial examples.
• Use the WDCNN to generate perturbed test instances that are then transferred to target models (e.g., LSTM, GRU, HDC) for evaluation.
• Encode raw sensor data into high-dimensional hypervectors (HVs) using HDC, enabling robust pattern recognition through simple vector operations like addition and dot product.
• Train HDC models using a simple learning rule based on vector similarity, avoiding backpropagation and enabling fast, energy-efficient training.
• Measure model resilience as the drop in classification accuracy before and after adversarial attacks, with lower degradation indicating higher resilience.
• Compare HDC’s performance against multiple deep learning models (LSTM, GRU, CRNN, etc.) across two datasets (CWRU bearing dataset, drive and fan end configurations).

Experimental results

Research questions

• RQ1 How does hyper-dimensional computing (HDC) compare to state-of-the-art deep learning models in resilience against black-box adversarial attacks in intelligent fault diagnosis?
• RQ2 To what extent does HDC reduce training time while maintaining or improving robustness in I-IoT fault diagnosis systems?
• RQ3 Can HDC maintain high classification accuracy under transferable adversarial attacks that exploit substitute models?
• RQ4 What is the average and maximum improvement in resiliency of HDC over deep learning baselines across different attack configurations and datasets?

Key findings

• HDC demonstrated up to 67.5% higher resilience than the best-performing deep learning model (GRU) on the fan end dataset under black-box adversarial attacks.
• HDC achieved a maximum resiliency improvement of 61.9% over the CGRU model on the drive end dataset, with average improvements exceeding 40% across all deep learning baselines.
• HDC reduced training time by up to 25.1× compared to the slowest deep learning model (LSTM), with a 23× speedup over the most resilient deep learning model (GRU).
• HDC maintained the highest classification accuracy under adversarial conditions, outperforming all deep learning models on both the drive end and fan end datasets.
• The average resiliency improvement of HDC over all deep learning models was 20% or higher, with the highest average improvement of 48.3% observed for CLSTM on the fan end.
• HDC’s lightweight training and robustness make it a viable solution for real-time, continuous retraining in I-IoT systems where data streams are dynamic and security is critical.