ABSTRACT
Neural networks have been widely applied in various fields, including drones and autonomous vehicles. The performance of neural networks determines their effectiveness, but reliability is equally important. Building on previous research on factors influencing neural network reliability, this study employed a fault injection framework to conduct experiments and further investigate the impact of layer types on network reliability. By conducting fault injection experiments, and observing the maximum bit error rate that different layers can tolerate, we can evaluate the reliability of each layer in the network model. Additionally, we introduced different types of layers into traditional neural network models and conducted experiments to further examine the reliability relationship among these layers within the same model. The results of the fault injection experiments indicate that the convolutional layer is the most susceptible to disruptions among the layers, with its reliability decreasing as the number of convolutional layers increases. On the other hand, the reliability of the fully connected layer improves with an increase in the number of layers.
- Chen, Y. H. 2016. Eyeriss: A Spatial Architecture for Energy-Efficient Dataflow for Convolutional Neural Networks. In The International Symposium on Computer Architecture (ISCA). DOI: 10.1109/ISCA.2016.40.Google ScholarDigital Library
- Krizhevsky, A. 2012. ImageNet Classification with Deep Convolutional Neural Networks. In Neural Information Processing Systems (NIPS). https://doi.org/10.1145/3065386Google ScholarDigital Library
- Chen, Z. 2019. BinFI: An Efficient Fault Injector for Safety-Critical Machine Learning Systems. In The International Conference for High Performance Computing, Networking, Storage and Analysis (SC). DOI: 10.1145/3295500.3356177Google ScholarDigital Library
- Fang, B. 2013. Towards Building Error Resilient GPGPU Applications. Blogs.ubc.ca https://api.semanticscholar.org/CorpusID:11740337Google Scholar
- Dos Santos, F. 2017. Evaluation and mitigation of soft-errors in neural network-based object detection in three GPU architectures. pp. 169-176. DOI: 10.1109/DSN-W.2017.47Google ScholarCross Ref
- Santos, F. F. d. 2019. Analyzing and Increasing the Reliability of Convolutional Neural Networks on GPUs. IEEE Transactions on Reliability, 68(2), 663-677. DOI: 10.1109/TR.2018.2878387.Google ScholarCross Ref
- Li, G., Pattabiraman, K. 2018. TensorFI: A Configurable Fault Injector for TensorFlow Applications. 2018 IEEE International Symposium on Software Reliability Engineering Workshops (ISSREW), pp. 313-320. DOI: 10.1109/ISSREW.2018.00024.Google ScholarCross Ref
- Reagen, B. 2018. Ares: A framework for quantifying the resilience of deep neural networks. IEEE Design Automation Conference (DAC), vol. 55, pp. 1-6. DOI:10.1109/DAC.2018.8465834.Google ScholarDigital Library
- Hoang, L.H. 2019. "FT-ClipAct: Resilience Analysis of Deep Neural Networks and Improving their Fault Tolerance using Clipped Activation." arXiv, DOI: 10.48550/arXiv.1912.00941.Google ScholarCross Ref
- Wei, J. 2020. Analyzing the impact of soft errors in VGG networks implemented on GPUs. Microelectronics Reliability, 110, 113648. DOI:10.1016/j.microrel.2020.113648.Google ScholarCross Ref
- Hari, S.K.S. 2017. Sassififi: an architecture level fault injection tool for GPU application resilience evaluation. In: Proceedings of the International Symposium on Performance Analysis of Systems and Software (ISPASS), California, USA, Oct 2017. DOI: 10.1109/ISPASS.2017.7975296Google ScholarCross Ref
- Mahmoud, A. 2020. PyTorchFI: A Runtime Perturbation Tool for DNNs. 2020 50th Annual IEEE/IFIP International Conference on Dependable Systems and Networks Workshops (DSN-W), Valencia, Spain, pp. 25-31. DOI: 10.1109/DSN-W50199.2020.00014.Google ScholarCross Ref
- Colucci, A. 2022. "enpheeph: A Fault Injection Framework for Spiking and Compressed Deep Neural Networks." 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Kyoto, Japan, pp. 5155-5162. DOI: 10.1109/IROS47612.2022.9982181.Google ScholarCross Ref
- Li, G. 2017. Understanding Error Propagation in Deep Learning Neural Network (DNN) Accelerators and Applications. SC17: International Conference for High Performance Computing, Networking, Storage and Analysis, Denver, CO, USA, pp. 1-12. Doi:10.1145/3126908.3126964Google Scholar
Index Terms
- Neural network reliability analysis based on fault injection
Recommendations
Fault Injection into VHDL Models: Experimental Validation of a Fault Tolerant Microcomputer System
EDCC-3: Proceedings of the Third European Dependable Computing Conference on Dependable ComputingThis work presents a campaign of fault injection to validate the dependability of a fault tolerant microcomputer system. The system is duplex with cold stand-by sparing, parity detection and a watchdog timer. The faults have been injected on a chip-...
Dependability Analysis of a High-Speed Network Using Software-Implemented Fault Injection and Simulated Fault Injection
This paper presents a dependability study of high-speed, switched Local Area Networks (LANs) using Myrinet as an example testbed (with theoretical speeds of 2.56 Gbps). The study uses results of two fault injection methods, simulated fault injection and ...
Combining Software-Implemented and Simulation-Based Fault Injection into a Single Fault Injection Method
FTCS '95: Proceedings of the Twenty-Fifth International Symposium on Fault-Tolerant ComputingAbstract: Fault/error injection has emerged as a valuable means for evaluating the dependability of a system. In particular, software-based techniques (which can be described as software-implemented and simulation-based techniques) have become very ...
Comments