[Paper Review] A Survey on Impact of Transient Faults on BNN Inference Accelerators

Navid Khoshavi, Connor Broyles+1 more|arXiv (Cornell University)|Apr 10, 2020

Advanced Neural Network Applications23 references13 citations

TL;DR

This paper investigates the impact of transient soft errors—specifically Single-Event Upsets (SEUs) and Multi-Bit Upsets (MBUs)—on Binarized Neural Network (BNN) inference accelerators using FPGA-based fault injection. It demonstrates that activation layers are highly vulnerable, with classification accuracy dropping by up to 76.7% in worst-case scenarios due to soft errors, especially under high fault accumulation.

ABSTRACT

Over past years, the philosophy for designing the artificial intelligence algorithms has significantly shifted towards automatically extracting the composable systems from massive data volumes. This paradigm shift has been expedited by the big data booming which enables us to easily access and analyze the highly large data sets. The most well-known class of big data analysis techniques is called deep learning. These models require significant computation power and extremely high memory accesses which necessitate the design of novel approaches to reduce the memory access and improve power efficiency while taking into account the development of domain-specific hardware accelerators to support the current and future data sizes and model structures.The current trends for designing application-specific integrated circuits barely consider the essential requirement for maintaining the complex neural network computation to be resilient in the presence of soft errors. The soft errors might strike either memory storage or combinational logic in the hardware accelerator that can affect the architectural behavior such that the precision of the results fall behind the minimum allowable correctness. In this study, we demonstrate that the impact of soft errors on a customized deep learning algorithm called Binarized Neural Network might cause drastic image misclassification. Our experimental results show that the accuracy of image classifier can drastically drop by 76.70% and 19.25% in lfcW1A1 and cnvW1A1 networks,respectively across CIFAR-10 and MNIST datasets during the fault injection for the worst-case scenarios

Motivation & Objective

• To analyze the real-world impact of transient soft errors on BNN inference accelerators, particularly in safety-critical applications like self-driving vehicles.
• To identify the most vulnerable components—specifically, layers, data types (weights, activations), and fault types (SEU, MBU)—within BNN accelerators.
• To evaluate how fault accumulation over time affects classification accuracy and system reliability in BNN accelerators.
• To develop and validate a realistic, FPGA-based fault injection framework to simulate soft errors in a way that reflects actual hardware behavior.

Proposed method

• Conducted fault injection experiments using a modified version of the FINN framework on FPGA-based BNN accelerators to simulate transient soft errors.
• Injected both SEUs (single-bit flips) and MBUs (multi-bit flips) across combinational logic and memory storage in BNN accelerators.
• Targeted different network layers—especially activation and weight layers—across two BNN architectures: lfcW1A1 and cnvW1A1.
• Measured classification accuracy degradation under varying fault counts (up to 100 faults) and fault distributions across time and space.
• Used CIFAR-10 and MNIST datasets to evaluate performance under realistic workloads and fault injection scenarios.
• Analyzed fault propagation by tracking how corrupted data in early layers (e.g., first convolutional layer) contaminates downstream computations.

Experimental results

Research questions

• RQ1 How do SEUs and MBUs affect the classification accuracy of BNN inference accelerators under realistic fault injection conditions?
• RQ2 Which layers and data types (weights, activations) in BNN accelerators are most vulnerable to soft errors?
• RQ3 How does the accumulation of soft errors over time impact the long-term reliability and accuracy of BNN inference?
• RQ4 What is the relative impact of MBUs compared to SEUs on BNN accelerator output accuracy?
• RQ5 How does the position of a layer in the network (e.g., early vs. late) influence its susceptibility to soft error-induced accuracy degradation?

Key findings

• The lfcW1A1 network experienced a worst-case accuracy drop of 76.7% (from 98.4% to 22.92%) when 100 SEUs were injected into the activation layer.
• The cnvW1A1 network showed a 19.25% accuracy degradation under worst-case SEU injection, while the lfcW1A1 network suffered a 76.7% drop.
• Activation layers were significantly more vulnerable than weight layers, with fault-induced accuracy drops becoming pronounced when more than 20 faults were accumulated.
• MBUs had a higher impact than SEUs due to the corruption of multiple bits and adjacent data, leading to greater error propagation and system-level degradation.
• The first layer in the network was the most vulnerable, with early-layer faults causing cascading errors that drastically reduced final classification accuracy.
• Even though average accuracy degradation across scenarios was moderate, the worst-case scenarios revealed extreme misclassification risks, especially in safety-critical applications.