Assessing the reliability of modern devices running CNN algorithms is a very difficult task. Actually, the complexity of the state-of-the-art devices makes exhaustive Fault Injection (FI) campaigns impractical and typically out of the computational capabilities. A possible solution consists of resorting to statistical FI campaigns that allow a reduction in the number of needed experiments by injecting only a carefully selected small part of it. Under specific hypothesis, statistical FIs guarantee an accurate picture of the problem, albeit selecting a reduced sample size. The main problems today are related to the choice of the sample size, the location of the faults, and the correct understanding of the statistical assumptions. The intent of this paper is twofold: first, we describe how to correctly specify statistical FIs for Convolutional Neural Networks; second, we propose a data analysis on the CNN parameters that drastically reduces the number of FIs needed to achieve statistically significant results without compromising the validity of the proposed method. The methodology is experimentally validated on two CNNs, ResNet-20 and MobileNetV2, and the results show that a statistical FI campaign on about 1.21% and 0.55% of the possible faults, provides very precise information of the CNN reliability. The statistical results have been confirmed by the exhaustive FI campaigns on the same cases of study.

In recent years, the reliability of Deep Neural Networks (DNN) has become the focus of an increasing number of research activities. In particular, researchers have focused on understanding how a DNN behaves when the underlying hardware is affected by a fault. This is a challenging task: slight changes in a network architecture can significantly impact how the network reacts to faults. There are several approaches to simulate the behaviour of a faulty network: the most accurate one is to perform low-level fault simulations. Nonetheless, this task is very time-consuming and costly to be implemented. Even though the injection time can be reduced by injecting faults at the application level, for sufficiently large networks, this time is still very high, requiring weeks to complete a single simulation. This work aims at providing a fast and accurate solution for injecting software-level faults in a DNN that is independent of its architecture and does not require any modification to its structure. For this reason, this paper introduces SCI-FI, a Smart, aCcurate and unIntrusive Fault-Injector. SCI-FI smartly reduces the fault injection time required for a complete fault simulation of the network by taking advantage of two fundamental mechanisms: Fault Dropping and Delayed Start. Experimental results from various ResNet, DenseNet and EfficientNet architectures targeting the CIFAR-10 and ImageNet datasets show that combining these techniques drastically reduces the simulation time, which can last up to 70% less.

The widespread use of artificial intelligence (AI)-based systems has raised several concerns about their deployment in safety-critical systems. Industry standards, such as ISO26262 for automotive, require detecting hardware faults during the mission of the device. Similarly, new standards are being released concerning the functional safety of AI systems (e.g., ISO/IEC CD TR 5469). Hardware solutions have been proposed for the infield testing of the hardware executing AI applications; however, when used in applications such as Convolutional Neural Networks (CNNs) in image processing tasks, their usage may increase the hardware cost and affect the application performances. In this paper, for the very first time, a methodology to develop high-quality test images, to be interleaved with the normal inference process of the CNN application is proposed. An Image Test Library (ITL) is developed targeting the on-line test of GPU functional units. The proposed approach does not require changing the actual CNN (thus incurring in costly memory loading operations) since it is able to exploit the actual CNN structure. Experimental results show that a 6-image ITL is able to achieve about 95% of stuck-at test coverage on the floating-point multipliers in a GPU. The obtained ITL requires a very low test application time, as well as a very low memory space for storing the test images and the golden test responses.

In the last decade, a huge effort has been spent on assessing the reliability of Convolutional Neural networks (CNNs), probably the most popular architecture for image classification tasks. However, modern Deep Neural Networks (DNNs) are rapidly overtaking CNNs, as state-of-the-art results for many tasks are achieved with the Transformers, innovative DNN models. Transformers' architecture introduces the concept of attention as an alternative to the classical convolution operation. The aim of this work is to propose a reliability analysis of the Swin Transformer, one of the most accurate DNN used for Image Classification, that greatly improves the results obtained by traditional CNNs. In particular, this paper shows that, similar to CNNs, Transformers are susceptible to single faults affecting weights and neurons. Furthermore, it is shown how output ranging, a well-known technique to reduce the impact of a fault in CNNs, is not as effective for the Transformer. The alternative solution proposed by this work is to introduce a ranging not only on the output, but also on the input and on the weight of the fully connected layers. Results show that, on average, the number of critical faults (i.e., that modify the network's output) affecting neurons decreases by a factor of 1.91, while for faults affecting the network's weights this value decreases by a factor of$1\cdot 10^{5}$.

Deep Learning, and in particular, Deep Neural Network (DNN) is nowadays widely used in many scenarios, including safety-critical applications such as autonomous driving. In this context, besides energy efficiency and performance, reliability plays a crucial role since a system failure can jeopardize human life. As with any other device, the reliability of hardware architectures running DNNs has to be evaluated, usually through costly fault injection campaigns. This paper explores approximation and fault resiliency of DNN accelerators. We propose to use approximate (AxC) arithmetic circuits to agilely emulate errors in hardware without performing fault injection on the DNN. To allow fast evaluation of AxC DNN, we developed an efficient GPU-based simulation framework. Further, we propose a fine-grain analysis of fault resiliency by examining fault propagation and masking in networks.

In the literature, it is argued that Deep Neural Networks (DNNs) possess a certain degree of robustness mainly for two reasons: their distributed and parallel architecture, and their redundancy introduced due to over provisioning. Indeed, they are made, as a matter of fact, of more neurons with respect to the minimal number required to perform the computations. It means that they could withstand errors in a bounded number of neurons and continue to function properly. However, it is also known that different neurons in DNNs have divergent fault tolerance capabilities. Neurons that contribute the least to the final prediction accuracy are less sensitive to errors. Conversely, the neurons that contribute most are considered critical because errors within them could seriously compromise the correct functionality of the DNN. This paper presents a software methodology based on a Triple Modular Redundancy technique, which aims at improving the overall reliability of the DNN, by selectively protecting a reduced set of critical neurons. Our findings indicate that the robustness of the DNNs can be enhanced, clearly, at the cost of a larger memory footprint and a small increase in the total execution time. The trade-offs as well as the improvements are discussed in the work by exploiting two DNN architectures: ResNet and DenseNet trained and tested on CIFAR-10.

This paper encompasses three contributions by industry professionals and university researchers. The contributions describe different trends in automotive products, including both manufacturing test and run-time reliability strategies. The subjects considered in this session deal with critical factors, from optimizing the final test before shipment to market to in-field reliability during operative life.

Deep Neural Networks (DNNs) are nowadays widely used in low-cost accelerators, characterized by limited computational resources. These models, and in particular DNNs for image classification, are becoming increasingly popular in safety-critical applications, where they are required to be highly reliable. Unfortunately, increasing DNNs reliability without computational overheads, which might not be affordable in low-power devices, is a non-trivial task. Our intuition is to detect network executions affected by faults as outliers with respect to the distribution of normal network’s output. To this purpose, we propose to exploit Open-Set Recognition (OSR) techniques to perform Fault Detection in an extremely low-cost manner. In particuar, we analyze the Maximum Logit Score (MLS), which is an established Open-Set Recognition technique, and compare it against other well-known OSR methods, namely OpenMax, energy-based outof-distribution detection and ODIN. Our experiments, performed on a ResNet-20 classifier trained on CIFAR-10 and SVHN datasets, demonstrate that MLS guarantees satisfactory detection performance while adding a negligible computational overhead. Most remarkably, MLS is extremely convenient to conFigure and deploy, as it does not require any modification or re-training of the existing network. A discussion of the advantages and limitations of the analysed solutions concludes the paper.