The Robustness of Modern Deep Learning Architectures against Single Event Upset Errors
The DOI matches its Crossref record; 36 of the 37 references it cites were found in a registry or an index.
Abstract
The following topics are dealt with: graph theory; parallel processing; graphics processing units; optimisation; field programmable gate arrays; learning (artificial intelligence); multiprocessing systems; data analysis; parallel architectures; microprocessor chips.
Overview
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
Citations
Cites
Deep Residual Learning for Image Recognition
He, K., Zhang, X., Ren, S., & Sun, J. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.90 · Cited by 235,474 · cited 3×
“In the case of this paper Microsoft’s 50 layer network is tested, although they also implemented a 152 layer network.”
InfluentialVery Deep Convolutional Networks for Large-Scale Image Recognition
Simonyan, K., & Zisserman, A. (2014) · arXiv · doi:10.48550/arxiv.1409.1556 · Cited by 114,148 · cited 2×
“A. VGG16 VGG16 was introduced in 2014 by the Visual Geometry Group at Oxford [27].”
MethodImageNet classification with deep convolutional neural networks
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2017) · Communications of the ACM · doi:10.1145/3065386 · Cited by 112,155 · cited 1×
“Neural networks’ greatest successes have been in computer vision tasks [1] [2] [3], but they have also been applied to tasks such as language processing [4], robotic control [5], and signal processing [6] [7].”
BackgroundImageNet: A large-scale hierarchical image database
Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., & Fei-Fei, L. (2009) · 2009 IEEE Conference on Computer Vision and Pattern Recognition · doi:10.1109/cvpr.2009.5206848 · Cited by 65,333 · cited 1×
“No training was needed because the models were instantiated with weights provided by the original researchers for classification against the ImageNet dataset [33].”
MethodGoing deeper with convolutions
Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., & Rabinovich, A. (2015) · 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2015.7298594 · Cited by 48,562
(2014) · Cited by 33,419
Rethinking the Inception Architecture for Computer Vision
Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., & Wojna, Z. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.308 · Cited by 32,138
Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift
Ioffe, S., & Szegedy, C. (2015) · arXiv · doi:10.48550/arxiv.1502.03167 · Cited by 27,818 · cited 1×
“Some differences are that ResNet50 uses a stride of 2 in the convolutional layers, does not use dropout during training, and utilizes batch normalization [30] between the convolutional layers.”
Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift
Ioffe, S. (2024) · arXiv (Cornell University) · doi:10.57702/o9raffed · Cited by 14,214
Explaining and Harnessing Adversarial Examples
Goodfellow, I. J., Shlens, J., & Szegedy, C. (2014) · arXiv · doi:10.48550/arxiv.1412.6572 · Cited by 10,381 · cited 1×
“An exciting research trend regarding the robustness of modern neural networks is the discovery of adversarial examples that can fool deep convolutional neural networks [23] [24].”
Background
Research works printed but not linked: 1
- 0, An On-device Deep Neural Network for Face Detection - Apple
Cited by
A survey on modeling and improving reliability of DNN algorithms and accelerators
Mittal, S. (2020) · Journal of Systems Architecture · doi:10.1016/j.sysarc.2019.101689 · Cited by 175
A Systematic Literature Review on Hardware Reliability Assessment Methods for Deep Neural Networks
Ahmadilivani, M. H., Taheri, M., Raik, J., Daneshtalab, M., & Jenihhin, M. (2023) · ACM Computing Surveys · doi:10.1145/3638242 · Cited by 132
“, PyTorch (used in [59][68][70][71][72][76][78]), Keras (used in [61][62][80][81]), TensorFlow (used in [66][79]), Caffe (used in [77]), DarkNet (used in [73][140][142]).”
MethodSoft errors in DNN accelerators: A comprehensive review
Ibrahim, Y., Wang, H., Liu, J., Wei, J., Chen, L., Rech, P., Adam, K., & Guo, G. (2020) · Microelectronics Reliability · doi:10.1016/j.microrel.2020.113969 · Cited by 102
Testability and Dependability of AI Hardware: Survey, Trends, Challenges, and Perspectives
Su, F., Liu, C., & Stratigopoulos, H. (2023) · IEEE Design and Test · doi:10.1109/mdat.2023.3241116 · Cited by 97
“The interested reader is referred to [70], [71], [72], [73], [74], [75], [76], [77], [78], [79], [80], [81], [82], [83], [84], and [85] for more software-based fault injection experiments studying the fault impact for different neural network models, data type representations, layer types, network sizes, pruned networ…”
BackgroundA Survey of Bit-Flip Attacks on Deep Neural Network and Corresponding Defense Methods
Qian, C., Zhang, M., Nie, Y., Lu, S., & Cao, H. (2023) · Electronics · doi:10.3390/electronics12040853 · Cited by 42
FTT-NAS: Discovering Fault-Tolerant Neural Architecture
Li, W., Ning, X., Ge, G., Chen, X., Wang, Y., & Yang, H. (2020) · 2020 25th Asia and South Pacific Design Automation Conference (ASP-DAC) · doi:10.1109/asp-dac47756.2020.9045324 · Cited by 34
Burel, S., Evans, A., & Anghel, L. (2022) · IEEE Transactions on Device and Materials Reliability · doi:10.1109/tdmr.2022.3159089 · Cited by 35
Wong, L. J., Clark, W. H., Flowers, B., Buehrer, R. M., Michaels, A. J., & Headley, W. C. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2010.00432 · Cited by 32
Self-citationDependable DNN Accelerator for Safety-Critical Systems: A Review on the Aging Perspective
Moghaddasi, I., Gorgin, S., & Lee, J. (2023) · IEEE Access · doi:10.1109/access.2023.3300376 · Cited by 30
“with lower cost and more resilience [32].”
FHDnn: communication efficient and robust federated learning for AIoT networks
Chandrasekaran, R., Ergun, K., Lee, J., Nanjunda, D., Kang, J., & Rosing, T. (2022) · ACM/IEEE Design Automation Conference (DAC) · doi:10.1145/3489517.3530394 · Cited by 28
“CNNs in particular are not robust to noise on weights as shown in [3].”
Wong, L. J., Clark, W. H., Flowers, B., Buehrer, R. M., Headley, W. C., & Michaels, A. J. (2021) · IEEE Open Journal of the Communications Society · doi:10.1109/ojcoms.2021.3112939 · Cited by 29
Self-citationGutiérrez‐Zaballa, J., Basterretxea, K., & Echanobe, J. (2024) · Journal of Systems Architecture · doi:10.1016/j.sysarc.2024.103242 · Cited by 29
Wagle, S. A., Harikrishnan, R., Sampe, J., Faseehuddin, M., & Ali, S. H. M. (2021) · Traitement du signal · doi:10.18280/ts.380609 · Cited by 24
“Their use of batch normalization or their use of shortcut connections is among the possible reasons for the strength of ResNet50 [14].”
BackgroundReliability Evaluation of Pruned Neural Networks against Errors on Parameters
Gao, Z., Wei, X., Zhang, H., Li, W., Ge, G., Wang, Y., & Reviriego, P. (2020) · Proceedings · doi:10.1109/dft50435.2020.9250812 · Cited by 24
“In [20], the reliability of three modern network structures VGG 16, ResNet50, and InceptionV3 is evaluated by flipping some bits of the main parameters, including weights, biases, and the batch normalization (BN) parameters.”
BackgroundMethodResultInfluentialReliability evaluation of FPGA based pruned neural networks
Gao, Z., Yao, Y., Wei, X., Yan, T., Zeng, S., Ge, G., Wang, Y., Ullah, A., & Reviriego, P. (2022) · Microelectronics Reliability · doi:10.1016/j.microrel.2022.114498 · Cited by 22
Systematic Reliability Evaluation of FPGA Implemented CNN Accelerators
Gao, Z., Gao, S., Yao, Y., Liu, Q., Zeng, S., Ge, G., Wang, Y., Ullah, A., & Reviriego, P. (2023) · IEEE Transactions on Device and Materials Reliability · doi:10.1109/tdmr.2023.3235767 · Cited by 21
Gao, Z., Yin, Z., Wang, J., Su, R., Deng, J., Liu, Q., Reviriego, P., Liu, S., & Lombardi, F. (2025) · IEEE Transactions on Nanotechnology · doi:10.1109/tnano.2025.3531721 · Cited by 18
“For example, [29] has investigated the effects of bit flips on single precision weights and found that the exponent bits are more vulnerable to errors; normalization layers are effective to reduce the impact of such errors.”
Selective Hardening of CNNs based on Layer Vulnerability Estimation
Bolchini, C., Cassano, L., Miele, A., & Nazzari, A. (2022) · Proceedings · doi:10.1109/dft56152.2022.9962339 · Cited by 21
“Other methodologies, like [21], [22], only support the corruption of the operator weights.”
MethodSyed, R. T., Zhao, Y., Chen, J., Andjelkovic, M. S., Ulbricht, M., & Krstic, M. D. (2024) · IEEE Access · doi:10.1109/access.2024.3392240 · Cited by 15
“Authors in [7] suggest a triple modular redundant reconfigurable DNN accelerator for data compression using an autoencoder model, while authors in [13] have performed reliability analysis of VGG16, ResNet50, and InceptionV3 against transient faults happening in the model’s parameter and summarized…”
MethodEvaluation and Mitigation of Weight-Related Single Event Upsets in a Convolutional Neural Network
Cai, Y., Cai, M., Wu, Y., Lü, J., Bian, Z., Liu, B., & Cui, S. (2024) · Electronics · doi:10.3390/electronics13071296 · Cited by 13
MOZART: Masking Outputs with Zeros for Architectural Robustness and Testing of DNN Accelerators
Burel, S., Evans, A., & Anghel, L. (2021) · IEEE International On-Line Testing Symposium (IOLTS) · doi:10.1109/iolts52814.2021.9486694 · Cited by 12
Gao, Z., Wang, J., Su, R., Reviriego, P., Liu, S., & Lombardi, F. (2023) · IEEE International Conference on Nanotechnology (IEEE-NANO) · doi:10.1109/nano58406.2023.10231204 · Cited by 9
“[19], [20] and [21] have focused on floating-point (FP) implementations, and it has been found that an error rate of 10-7 dramatically degrade the classification accuracy; also, errors on exponent bits are critical, because they reduce accuracy the most, so degrading dependable operations.”
BackgroundStudy of Sensitivity to Weight Perturbation for Convolution Neural Network
Xiang, L., Zeng, X., Niu, Y., & Liu, Y. (2019) · IEEE Access · doi:10.1109/access.2019.2926768 · Cited by 9
VR-ZYCAP: A Versatile Resourse-Level ICAP Controller for ZYNQ SOC
Sultana, B., Ullah, A., Malik, A. A., Zahir, A., Reviriego, P., Muslim, F. B., Ullah, N., & Ahmad, W. (2021) · Electronics · doi:10.3390/electronics10080899 · Cited by 9
FTT-NAS: Discovering Fault-Tolerant Neural Architecture.Preprint
Ning, X., Ge, G., Li, W., Zhu, Z., Zheng, Y., Chen, X., Gao, Z., Wang, Y., & Yang, H. (2020) · arXiv (Cornell University) · Cited by 9
Dependability and Protection of Transformer Models Against Soft Errors on Text Embeddings
Gao, Z., Liu, S., Reviriego, P., Liu, S., & Lombardi, F. (2024) · IEEE Transactions on Device and Materials Reliability · doi:10.1109/tdmr.2024.3478753 · Cited by 8
“The dependability of convolutional DNN models has been well studied during the last ten years [18], [19], [20], [21], and researchers have also started recently to study the dependability of Transformers.”
Towards Reliable AI Applications via Algorithm-Based Fault Tolerance on NVDLA
Saniç, M. T., Guo, C., Leng, J., Guo, M., & Ma, W. (2022) · International Conference on Mobility, Sensing and Networking (MSN) · doi:10.1109/msn57253.2022.00120 · Cited by 8
“The resilience of DNN architectures is examined with bit flip attacks [20].”
BackgroundMethodRojas, E., Pérez, D., & Meneses, E. (2024) · Journal of Parallel and Distributed Computing · doi:10.1016/j.jpdc.2024.104879 · Cited by 5
Dependability Evaluation of Stable Diffusion with Soft Errors on the Model Parameters
Gao, Z., Yuan, L., Reviriego, P., Liu, S., & Lombardi, F. (2024) · IEEE International Conference on Nanotechnology (IEEE-NANO) · doi:10.1109/nano61778.2024.10628863 · Cited by 5
Exploring the Effects of Silent Data Corruption in Distributed Deep Learning Training
Rojas, E., Perez, D., & Meneses, E. (2022) · International Symposium on Computer Architecture and High Performance Computing (SBAC-PAD) · doi:10.1109/sbac-pad55451.2022.00013 · Cited by 5
“A similar approach is presented in [13] in which errors are evaluated after training is complete in the object classification stage.”
MethodOperating Conversational Large Language Models (LLMs)in the Presenceof Errors
Gao, Z., Deng, J., Reviriego, P., Liu, S., Pozo, A., & Lombardi, F. (2024) · IEEE Nanotechnology Magazine · doi:10.1109/mnano.2024.3513112 · Cited by 3
“[14].”
Resilience Improvements for Space-Based Radio Frequency Machine Learning
Wong, L. J., Altland, E., Detwiler, J., Fermin, P., Kuzin, J. M., Moeliono, N., Abdalla, A. S., Headley, W. C., & Michaels, A. J. (2020) · International Symposium on Networks, Computers and Communications (ISNCC) · doi:10.1109/isncc49221.2020.9297212 · Cited by 2
“nets (CNNs), which is a common structure to each of the RF, computer vision, and image recognition modalities, and ask the fundamental question: “How can the performance of an RFML neural net be best protected in the presence of radiation-induced single event upsets?” Prior work in this area [16]–[18] analyzed the rat…”
BackgroundSelf-citationFault injection analysis of Real NVP normalising flow model for satellite anomaly detection
Greco, G., Cena, C., Albertin, U., Martini, M., & Chiaberge, M. (2025) · International Joint Conference on Neural Networks (IJCNN) · doi:10.1109/ijcnn64981.2025.11227924 · Cited by 2
“…costs, better controllability, and easier deployment for developers, the commonly adopted fault model (typically bit flip) [13, 14, 15, 16] may be accurate for main memory structures but less realistic for faults within computing cores or control logic, where the programmer has limited influence.”
Fast and Accurate Error Simulation for CNNs Against Soft Errors
Bolchini, C., Cassano, L., Miele, A., & Toschi, A. (2023) · IEEE Transactions on Computers · doi:10.1109/tc.2022.3184274 · Cited by 1
“[19], [32] introduce two tools developed in Keras and Darknet, supporting only the corruption of the operator weights in the CNN.”
BackgroundMethodButts, C., Jha, R., Messay-Kebede, T., & Kapp, D. (2022) · IEEE Transactions on Computers · doi:10.1109/tc.2022.3204230 · Cited by 1
“In [19] the authors show that ANN architectures are somewhat resilient to SEU errors by introducing small amounts of error to randomly selected weights within the network.”
BackgroundOn the Impact of Soft Errors on Transformers Embeddings
Gao, Z., Liu, S., Reviriego, P., Liu, S., & Lombardi, F. (2024) · IEEE International Conference on Nanotechnology (IEEE-NANO) · doi:10.1109/nano61778.2024.10628838 · Cited by 1
Evaluation and Mitigation of Single Event Upset on the Accuracy of Neural Network Accelerators
Huang, Y., Liu, K., Zhu, W., & Zhang, N. (2024) · IEEE Information Technology, Networking, Electronic and Automation Control Conference (ITNEC) · doi:10.1109/itnec60942.2024.10733146 · Cited by 1
“Arechiga et al. [13] studied three DNNs for their robustness against DRAM errors.”
BackgroundA Fault Injection Method for Evaluating Single Event Soft Errors in Deep Neural Networks
Lyu, H., Mei, B., Xue, P., Wang, B., Mo, R., Yuan, Q., & Gao, W. (2025) · International Conference on Radiation Effects of Electronic Devices (ICREED) · doi:10.1109/icreed65908.2025.11036247 · Cited by 1
“Among the comprehensively studied architectures are VGG16, AlexNet, ResNet50, Inception, MobileNet, and LeNet, which have been systematically analyzed in existing literature [21,22].”
Maffazioli, U., Benevenuti, F., Medina, N. H., Added, N., Aguiar, V. A. P., Alberton, S. G. P. N., Schervenin, J. V., Pereira, M. S., Filho, A. C. B., Azambuja, J. R., et al. (2026) · IEEE Transactions on Nuclear Science · doi:10.1109/tns.2026.3691993 · Cited by 1
“Prior work has shown that bit flips in trained parameters cause sharp bimodal accuracy degradation, with as few as tens of simultaneous errors driving modern architectures to near-random outputs [32].”
Architectures to Ensure the Functional Safety of Neural Network based Systems
Burel, S. (2022) · theses.fr (ABES) · doi:10.70675/08b0620bzfafaz401fz8e91z6856bff62a00 · Cited by 1
Geometric algorithms for predicting resilience and recovering damage in neural networksPreprint
Raghavan, G., Li, J., & Thomson, M. W. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2005.11603
“Errors in dynamic access memory can occur due to malicious attacks (the RowHammer), but a particular focus has been on errors induced by high energy particles [21] that occur at surprising rates [22].”
BackgroundDependability of Alternative Computing Paradigms for Machine Learning: hype or hope?
Bolchini, C., Bosio, A., Cassano, L., Deveautour, B., Natale, G. D., Miele, A., O’Connor, I., & Vatajelu, E. I. (2022) · IEEE International Symposium on Design and Diagnostics of Electronic Circuits & Systems (DDECS) · doi:10.1109/ddecs54261.2022.9770138 · Cited by 1
“Several papers (e.g. [50], [51]) analyzed how faults corrupting the weights of the convolutional layers affect the final output of the CNN.”
BackgroundComplex Network Theory-based Deep Neural Network Degradation Analysis in the Context of Bit Attack
Shao, Q., Yang, S., Zhou, Y., Lu, D. Z., & Yao, Q. (2024) · IEEE International Conference on Software Quality, Reliability and Security Companion (QRS-C) · doi:10.1109/qrs-c63300.2024.00040
“Studies such [11] as have investigated the effects of single-bit and multi-bit perturbations on various models, including multilayer perceptron, pure convolutional networks, VGG16, and ResNet50, within floating-point storage formats.”
BackgroundMa, M., Chi, Y., Chen, J., & Liang, B. (2025) · International Conference on Radiation Effects of Electronic Devices (ICREED) · doi:10.1109/icreed65908.2025.11036189
“The sensitivity of network parameters to SEU is related to the function of the layer and their position within the neural network structure [4][5][6] .”
Anomaly Detection Reliability in Consumer Electronics Manufacturing
Pai, C.-W., Hsueh, H.-W., & Hsu, S.-H. (2025) · IEEE Transactions on Consumer Electronics · doi:10.1109/tce.2025.3626738
“Arechiga et al. show the importance of reliability by evaluating three representative convolution neural network architectures (CNNs), VGG16, ResNet50, and Inception V3 [20].”
Fault-tolerant CNN accelerator with reconfigurable capabilities
Syed, R. T. (2025) · publish.UP (University of Potsdam) · doi:10.25932/publishup-68754
Bouvier, M. (2021) · theses.fr (ABES) · doi:10.70675/39a95f62z4a96z486azbdc1z8ebf1379c1b1 · Cited by 3
bashi, F. N. K., Khairullah, S. S., & Dawwd, S. A. (2026) · Digital · doi:10.3390/digital6030054
“More sensitive components such as exponent bits, sign bits, and important neurons are more susceptible, and memory buffer faults are more dangerous than datapath faults due to data reuse [58,59], which enables a single fault to harm numerous computations [44,60].”
Dependability and Error Detection for the Stable Diffusion Generative Model
Gao, Z., Yuan, L., Conde, J., Reviriego, P., Alonso, Á., Liu, S., & Lombardi, F. (2026) · IEEE Transactions on Reliability · doi:10.1109/tr.2026.3734708
“…a focal point of prior research, particularly for CNNs. Studies have demonstrated that an SEU rate as low as 10 − 7 can cause substantial performance degradation in CNNs with floating-point weights, while fixed-point representations exhibit greater resilience [13], [14], [15], [16], [17], [18], [19].”
Metadata
- Authors
- Arechiga, A. P., & Michaels, A. J.
- ORCID
- Michaels, Alan J.
- Venue
- 2018 IEEE High Performance extreme Computing Conference (HPEC)
- Publisher
- IEEE
- Published
- 2018-09
- Pages
- 1-6
- OpenAlex
- W2902083778
- Cited by
- 47
- References
- 37
- Citations by year
- 2026: 2 · 2025: 5 · 2024: 10 · 2023: 7 · 2022: 9 · 2021: 5 · 2020: 7 · 2019: 2
- FWCI
- 5.58
- Type
- conference paper
- Language
- en
- Open access
- closed
- Locations
- IEEE High Performance Extreme Computing Conference (publishedVersion)
- Topics
- Physical Sciences › Engineering › Electrical and Electronic Engineering › Radiation Effects in Electronics
- Physical Sciences › Computer Science › Hardware and Architecture › Parallel Computing and Optimization Techniques
- Physical Sciences › Engineering › Electrical and Electronic Engineering › Advanced Memory and Neural Computing
- Keywords
- single event upset, deep learning robustness, hardware errors, fault tolerance, neural network reliability
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.