Characterizing Deep Learning Neural Network Failures Between Algorithmic Inaccuracy and Transient Hardware Faults

Laskar, Sabuj; Rahman, Md. Hasanur; Zhang, Bohan; Li, Guanpeng

doi:10.1109/prdc55274.2022.00020

Cited by 2 publications

(4 citation statements)

References 39 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In delineating the distinctions between this study and our earlier work by Laskar et al [21], we note that our prior research did not delve into the intricate dynamics of how diverse supergroup formations and the varying vulnerability of distinct DNN regions contribute to the probability of SCMs. This article, serving as a substantially revised and expanded edition of our previous work, offers a more profound comprehension of the multifaceted factors influencing the likelihood of SCMs across different DNN models.…”

Section: Introductionmentioning

confidence: 67%

“…In Section 5.1, we discuss our method of forming supergroups from various image classes of the CIFAR-100 and Ima-geNet datasets. This is the default supergroup formation that we used in our previous study [21]. However, in our previous study by Laskar et al [21], although we carefully analysed and created the formation of the supergroups in the context of safety critical behaviour, it is important to note that the configuration of these groups may vary depending on the specific requirements of the end-user.…”

Section: Methods For Forming Different Supergroup Setsmentioning

confidence: 99%

“…This is the default supergroup formation that we used in our previous study [21]. However, in our previous study by Laskar et al [21], although we carefully analysed and created the formation of the supergroups in the context of safety critical behaviour, it is important to note that the configuration of these groups may vary depending on the specific requirements of the end-user. Therefore, researchers are further interested to find the answer of the following research question in finer granularity: Would the probabilities of SCM and non-SCM remain consistent with the findings in our previous study by Laskar et al [21], especially when we alter the supergroup formation while still prioritizing safety critical perspectives?…”

Section: Methods For Forming Different Supergroup Setsmentioning

confidence: 99%

“…In our previous study [21], we answered the above question by characterizing DNN misclassifications caused by DNN intrinsic algorithmic inaccuracy as well as transient hardware faults in the context of safety-critical applications such as AVs. Our hypothesis posited that misclassifications arising from transient hardware faults pose a greater safety concern than those caused by intrinsic algorithmic inaccuracies.…”

mentioning

confidence: 99%

See 3 more Smart Citations

Investigating the impact of transient hardware faults on deep learning neural network inference

Rahman,

Laskar,

2024

Software Testing Verif & Rel

Self Cite

View full text Add to dashboard Cite

SummarySafety‐critical applications, such as autonomous vehicles, healthcare, and space applications, have witnessed widespread deployment of deep neural networks (DNNs). Inherent algorithmic inaccuracies have consistently been a prevalent cause of misclassifications, even in modern DNNs. Simultaneously, with an ongoing effort to minimize the footprint of contemporary chip design, there is a continual rise in the likelihood of transient hardware faults in deployed DNN models. Consequently, researchers have wondered the extent to which these faults contribute to DNN misclassifications compared to algorithmic inaccuracies. This article delves into the impact of DNN misclassifications caused by transient hardware faults and intrinsic algorithmic inaccuracies in safety‐critical applications. Initially, we enhance a cutting‐edge fault injector, TensorFI, for TensorFlow applications to facilitate fault injections on modern DNN non‐sequential models in a scalable manner. Subsequently, we analyse the DNN‐inferred outcomes based on our defined safety‐critical metrics. Finally, we conduct extensive fault injection experiments and a comprehensive analysis to achieve the following objectives: (1) investigate the impact of different target class groupings on DNN failures and (2) pinpoint the most vulnerable bit locations within tensors, as well as DNN layers accountable for the majority of safety‐critical misclassifications. Our findings regarding different grouping formations reveal that failures induced by transient hardware faults can have a substantially greater impact (with a probability up to 4 higher) on safety‐critical applications compared to those resulting from algorithmic inaccuracies. Additionally, our investigation demonstrates that higher order bit positions in tensors, as well as initial and final layers of DNNs, necessitate prioritized protection compared to other regions.

show abstract

Section: Introductionmentioning

confidence: 67%

Section: Methods For Forming Different Supergroup Setsmentioning

confidence: 99%