Evaluating Single Event Upsets in Deep Neural Networks for Semantic Segmentation: an embedded system perspective

TL;DR

This work investigates the robustness of encoder-decoder semantic segmentation DNNs to single event upsets (SEUs) when deployed on edge hardware. It conducts a comprehensive layer-by-layer and bit-level fault-injection study on a U‑Net model, comparing 32-bit FP32 and 8-bit INT quantized representations, across unpruned and pruned variants. The authors show that BN gamma and early encoder weights are the most vulnerable, while deep base layers are relatively inert; quantization tends to improve resilience by limiting magnitude growth and removing NaN-capable states. Crucially, the paper proposes memory- and computation-free mitigation via parameterization-based protection, alongside additional low-overhead techniques (sparse pruning, batch norm reconditioning, cross-layer equalization, and bias absorption) to enhance edge reliability without increasing resource use. The results inform practical design rules for robust embedded DNNs, enabling safer deployment in safety-critical domains such as ADAS/ADS and aerospace.

Abstract

As the deployment of artifical intelligence (AI) algorithms at edge devices becomes increasingly prevalent, enhancing the robustness and reliability of autonomous AI-based perception and decision systems is becoming as relevant as precision and performance, especially in applications areas considered safety-critical such as autonomous driving and aerospace. This paper delves into the robustness assessment in embedded Deep Neural Networks (DNNs), particularly focusing on the impact of parameter perturbations produced by single event upsets (SEUs) on convolutional neural networks (CNN) for image semantic segmentation. By scrutinizing the layer-by-layer and bit-by-bit sensitivity of various encoder-decoder models to soft errors, this study thoroughly investigates the vulnerability of segmentation DNNs to SEUs and evaluates the consequences of techniques like model pruning and parameter quantization on the robustness of compressed models aimed at embedded implementations. The findings offer valuable insights into the mechanisms underlying SEU-induced failures that allow for evaluating the robustness of DNNs once trained in advance. Moreover, based on the collected data, we propose a set of practical lightweight error mitigation techniques with no memory or computational cost suitable for resource-constrained deployments. The code used to perform the fault injection (FI) campaign is available at https://github.com/jonGuti13/TensorFI2 , while the code to implement proposed techniques is available at https://github.com/jonGuti13/parameterProtection .

Figures (37)

• Figure 1: A sample segmented image of HSI-Drive v2.0.
• Figure 2: Architecture of the unpruned unfolded unquantized U-Net model which has been used as reference.
• Figure 3: Single-precision floating-point format (IEEE 754 format).
• Figure 4: Bit-flip error rate on the original unpruned model.
• Figure 5: Bit-flip error rate on the original pruned model.