Robust Computing for Machine Learning-Based Systems

Muhammad Abdullah Hanif, Faiq Khalid, Rachmad Vidya Wicaksana Putra, Mohammad Taghi Teimoori, Florian Kriebel, Jeff Zhang, Kang Liu, Semeen Rehman, Theocharis Theocharides, Alessandro Artusi, Siddharth Garg, Muhammad ShafiqueEmbedded 2020 · DOI · OpenAlex

Abstract

The drive for automation and constant monitoring has led to rapid development in the field of Machine Learning (ML). The high accuracy offered by the state-of-the-art ML algorithms like Deep Neural Networks (DNNs) has paved the way for these algorithms to being used even in the emerging safety-critical applications, e.g., autonomous driving and smart healthcare. However, these applications require assurance about the functionality of the underlying systems/algorithms. Therefore, the robustness of these ML algorithms to different reliability and security threats has to be thoroughly studied and mechanisms/methodologies have to be designed which result in increased inherent resilience of these ML algorithms. Since traditional reliability measures like spatial and temporal redundancy are costly, they may not be feasible for DNN-based ML systems which are already super computer and memory intensive. Hence, new robustness methods for ML systems are required. Towards this, in this chapter, we present our analyses illustrating the impact of different reliability and security vulnerabilities on the accuracy of DNNs. We also discuss techniques that can be employed to design ML algorithms such that they are inherently resilient to reliability and security threats. Towards the end, the chapter provides open research challenges and further research opportunities.

Introduction

As Deep Neural Networks (DNNs) transition into safety-critical domains like autonomous driving and medical diagnostics, their massive computational demands collide with the necessity for extreme reliability. Modern AI systems often push hardware to its limits through aggressive voltage scaling or specialized accelerators to meet real-time constraints on edge devices. However, this optimization creates significant vulnerabilities. This paper explores the development of "inherently resilient" systems capable of surviving hardware faults and malicious attacks without the prohibitive power costs associated with traditional redundancy.

Vulnerabilities in ML-Based Systems

Robustness in ML systems is threatened by two primary categories of vulnerabilities: hardware-level reliability issues and software-level security threats. Hardware vulnerabilities include aging effects, "soft errors" (random bit-flips caused by cosmic rays), and timing faults induced by undervolting. Security threats involve adversarial perturbations—invisible noise designed to force misclassifications—and "backdoors" or hidden triggers within the model.

Overview of software and hardware-level vulnerabilities in ML-based systems.

Traditional solutions, such as Triple Modular Redundancy (TMR), are often too energy-expensive for power-constrained AI hardware. Consequently, a cross-layer approach is required to treat robustness as a primary design objective rather than an afterthought.

Proposed Robustness Methodologies

To address these challenges, several novel strategies are proposed to mitigate vulnerabilities while maintaining efficiency:

Taxonomy of reliability and security techniques for robust deep learning.
  • Fault-Aware Pruning (FAP & FAP+T): This technique identifies permanently broken hardware units (such as Multiply-Accumulate units) and prunes the weights associated with them. This effectively treats hardware faults as a form of "dropout" regularization.
  • Thundervolt Framework: This framework enables aggressive voltage underscaling. It utilizes TE-Drop, which detects timing errors but skips the erroneous operation instead of stalling for re-execution, leveraging the inherent error-resilience of neural networks.
  • Per-Layer Voltage Underscaling: By recognizing that different layers in a DNN have varying sensitivities, the system dynamically adjusts voltage on a layer-by-layer basis to maximize energy savings without compromising accuracy.
  • GAN-based Defense: Generative Adversarial Networks (GANs) are employed to model and remove environmental noise (e.g., haze or fog) that could be exploited by attackers to confuse vision-based systems.

Key Insights and Experimental Results

Experimental evaluation using standard benchmarks like MNIST, TIMIT, and ImageNet revealed several critical insights:

  1. The "Dropout" Parallel: Hardware faults are not necessarily fatal. Because DNNs are over-parameterized, faulty components can be bypassed similarly to how "dead" neurons are handled during training.
  2. Bit-Flip Asymmetry: Research indicates that 0-to-1 bit-flips in high-significance bits are significantly more damaging than 1-to-0 flips, suggesting that hardware protection (like ECC) should be prioritized for specific bit positions.
  3. Layer-Specific Sensitivity: Some layers can tolerate massive timing errors with zero impact on final accuracy, while others are highly brittle. Exploiting this allows for more aggressive power savings than uniform voltage settings.

In testing on the TIMIT speech dataset, baseline accuracy plummeted from 74% to ~40% with only four faulty MAC units. However, using FAP+T, systems maintained near-perfect accuracy even when 50% of hardware units were faulty. Furthermore, the Thundervolt dynamic scaling approach significantly outperformed static scaling, maintaining Top-5 accuracy on ImageNet at much lower energy levels.

Conclusion and Future Work

While these methodologies provide a path toward robust ML, challenges remain. The FAP+T method currently requires per-chip retraining, which is difficult to scale for mass manufacturing. Additionally, there is a need for better interpretability to understand why certain neurons are more critical than others. As architectures evolve, the "arms race" between attackers and defenders will continue, necessitating constant innovation in cross-layer robust computing design.

Summary updated

Abstract

Abstract The drive for automation and constant monitoring has led to rapid development in the field of Machine Learning (ML). The high accuracy offered by the state-of-the-art ML algorithms like Deep Neural Networks (DNNs) has paved the way for these algorithms to being used even in the emerging safety-critical applications, e.g., autonomous driving and smart healthcare. However, these applications require assurance about the functionality of the underlying systems/algorithms. Therefore, the robustness of these ML algorithms to different reliability and security threats has to be thoroughly studied and mechanisms/methodologies have to be designed which result in increased inherent resilience of these ML algorithms. Since traditional reliability measures like spatial and temporal redundancy are costly, they may not be feasible for DNN-based ML systems which are already super computer and memory intensive. Hence, new robustness methods for ML systems are required. Towards this, in this chapter, we present our analyses illustrating the impact of different reliability and security vulnerabilities on the accuracy of DNNs. We also discuss techniques that can be employed to design ML algorithms such that they are inherently resilient to reliability and security threats. Towards the end, the chapter provides open research challenges and further research opportunities.
BibTeX
@inproceedings{hanif2020robust,
  author = {Muhammad Abdullah Hanif and Faiq Khalid and Rachmad Vidya Wicaksana Putra and Mohammad Taghi Teimoori and Florian Kriebel and Jeff Zhang and Kang Liu and Semeen Rehman and Theocharis Theocharides and Alessandro Artusi and Siddharth Garg and Muhammad Shafique},
  title = {{Robust Computing for Machine Learning-Based Systems}},
  booktitle = {Embedded},
  year = {2020},
  doi = {10.1007/978-3-030-52017-5\_20},
}
Figure