Split Computing has emerged as a promising paradigm for deploying Deep Neural Networks in Edge and Inter-net of Things systems, enabling inference tasks to be distributed between resource-constrained edge devices and cloud servers. This approach is particularly attractive for autonomous systems, where security and reliability may be critical. However, interme-diate feature maps transmitted between devices are vulnerable to corruption, which may result from intentional adversarial attacks or unintentional hardware faults. Distinguishing whether corruption originates from an external adversary or an inherent system fault is crucial for implementing appropriate counter-measures-reinforcing security mechanisms against attacks or improving system reliability to mitigate the effects of hardware-related faults. To the best of our knowledge, this work is the first to propose a machine learning-based classification mechanism capable of differentiating adversarial attacks from hardware defects in Split Computing systems. The proposed approach analyzes the intermediate feature maps transmitted from the edge device to the server, classifying the source of corruption to guide appropriate responses. Experimental results demonstrate that one of the proposed classifiers can distinguish between intentional and unintentional feature map corruptions with an accuracy of 93.91 %.

Non-destructive inspection of penetrable vertical structures demands innovative investigation paradigms and ad-hoc tools to support timely and efficient maintenance interventions. Although Ground Penetrating Radar (GPR) is an established technology, the inspection of large vertical structures with classical ground-based systems remains challenging when there is limited accessibility, the equipment is difficult to maneuver, the operator’s safety is placed at risk, and the time to perform the survey becomes long. A promising solution to address these limitations suggests the use of mini drones equipped with GPR systems. However, this possibility introduces new challenges in both system design and data processing. A practical approach for drone-based GPR monitoring involves adapting a traditional contact GPR system, combined with appropriate data processing techniques, to perform contactless surveys. This study delves deeper into this approach by comparing the imaging performance achieved when the same ground-coupled GPR system paired with a microwave tomography-based data processing is exploited for contact and contactless surveys. Experimental results related to the inspection of a reinforced concrete wall are presented and discussed.

Ground penetrating radar imaging is mostly tackled by resorting to approximate linear inversion algorithms that provide only qualitative maps of the probed scene in terms of location and approximate geometry of the buried anomalies. Deep learning techniques have recently been proposed to retrieve quantitative information as cost-effective alternatives to nonlinear inversion approaches. Indeed, deep neural networks can effectively learn to map the input data into spatial maps describing the electromagnetic properties of the targets. In this frame, the present paper considers the popular U-NET topology for performing quantitative subsurface imaging. Two different training strategies differing for the type of input data are examined and compared. The first one assumes the radargram in the time domain as the network input; differently, in the second one, the network takes in input a microwave tomographic image of the subsurface scene. Numerical results based on full-wave synthetic data and some experimental tests are reported to assess and compare the reconstruction performance of both training schemes.

1The reliability of Neural Networks has gained significant attention, prompting efforts to develop SW-based hardening techniques for safety-critical scenarios. However, evaluating hardening techniques using application-level fault injection (FI) strategies, which are commonly hardware-agnostic, may yield misleading results. This study for the first time compares two FI approaches (at the application level (APP) and instruction level (ISA)) to evaluate deep neural network SW hardening strategies. Results show that injecting permanent faults at ISA (a more detailed abstraction level than APP) changes completely the ranking of SW hardening techniques, in terms of both reliability and accuracy. These results highlight the relevance of using an adequate analysis abstraction for evaluating such techniques.

Artificial intelligence is becoming increasingly popular for IoT applications in safety-critical fields (e.g., autonomous systems and biomedical, robots). Unfortunately, the inference’s workload process alone increases as the model size grows. To meet the computational power limitations of mobile devices running IoT applications, modern services sometimes resort to the Split Computing paradigm. Split Computing divides the inference process of a Neural Network into Head and Tail for their execution in a mobile device and a server, respectively, which also allows the reduction of the overall IoT device’s computational cost. Nonetheless, Split Computing can be used in safety-critical fields where reliability is crucial, especially when mobile devices have computational and cost restrictions. This paper introduces hardening techniques acting on the software to mitigate the effects of hardware faults on Split Computing models. The proposed hardening techniques consist of i) a bounded activation function whose thresholds are refined by training, and ii) a per-channel bounding of the bottleneck quantization of the split points. To quantitatively assess their effectiveness, we resorted to two different split configurations of a model for image classification. In addition, we considered a Split Computing model for object detection. Our findings indicate that the proposed approaches effectively reduces fault effects by $\mathbf{3. 5 \%}$ for image classifiers and $5.73 \%$ for object detectors when compared with other hardening approaches for general DNNs.

The adoption of Deep Neural Networks (DNNs) in several domains allows for increased effectiveness in applications that deal with massive data-intensive and complex data inputs. When employed in safety-critical scenarios, such as automotive, aerospace, healthcare, and autonomous robotics, assessing the DNNs' reliability and functional safety is crucial to ensure their correct in-field operation, even in the presence of hardware faults. However, the system complexity and the massive amounts of data to be processed by DNNs prevent the effective adoption of traditional strategies for reliability characterization and for identifying the most fault-sensitive structures. Accurate fault assessment strategies usually require unacceptable computational power and large evaluation times. On the other hand, faster strategies commonly lack accuracy in correctly representing system faults. Consequently, it is necessary to develop effective strategies that trade-off between performance and accuracy. This work analyses three reliability assessment strategies for deep neural networks and their underlying hardware, highlighting the main solutions and challenges in terms of evaluation performance and fault characterization accuracy. We overview different solutions to evaluate the hardware accelerators implementing DNNs at three abstraction levels:$i$) by physically injecting faults on a GPU running DNNs, ii) by performing microarchitectural characterization of GPUs to develop application-accurate error models, and iii) by using structure-aware cross-layer error modeling on DNN hardware accelerators. Our experimental results indicate that accurate error representation requires structural features from the targeted hardware.

This letter deals with full 3-D imaging by contactless multimonostatic ground penetrating radar (GPR) data by focusing on the effect of the measurement configuration on reconstruction performance. The imaging is faced as a linear inverse scattering problem and the truncated singular value decomposition (TSVD) is adopted to obtain the regularized solution. A criterion to determine a suitable spacing among the measurement points is provided and the effect of the data under sampling on the imaging results is also investigated. Numerical results based on the system point spread function are provided to verify the validity of the proposed criterion for the estimation of the nonredundant measurement spacing. Finally, reconstruction results relevant to a laboratory controlled experimental test validate the imaging approach as well as the derived data sampling criterion.

This letter addresses the classical problem of estimating the achievable resolution in terms of the configuration parameters and features of the background media in microwave imaging problems. In particular, we focus on the 2-D scalar case and a homogeneous medium, while data are collected in the near field by a multimonostatic multifrequency configuration. An analytical formula is derived to estimate the transverse resolution. The formula is determined by resorting to the evaluation of the point-spread function (PSF) using the weighted adjoint method. The effect of the configuration parameters on the transverse resolution is investigated. Numerical results confirm the spatially varying behavior of the transverse resolution and the accuracy of the proposed formula by assessing the capability to distinguish two nearby objects.

Radar imaging algorithms generally exploit linear models of the electromagnetic scattering phenomenon. This assumption leads to qualitative and computationally effective data inversion schemes, which only account for direct scattering from targets, whereas multipath signal contributions are neglected. As a result, multipath ghosts, i.e. false targets reconstructed at positions where no real target exists, affect the radar images thus preventing a reliable interpretation of the observed scene. This paper proposes a fully data-driven deep learning approach based on a convolutional neural network and microwave tomography to face this challenge. The approach achieves multipath ghost suppression for the case of small targets in terms of probing wavelength. In the proposed training scheme, the tomographic image affected by ghosts represents the input of the network while a ghost-free reconstruction is the output. Numerical simulations addressing the detection of metallic rebars via ground penetrating radar are presented. As shown, the proposed ghost removal strategy is effective and robust to variations of the scenario parameters on which the network is trained. Finally, an experimental validation shows the effectiveness of the proposed strategy even in operative conditions.

This paper presents an imaging approach for Multiple Input Multiple Output Ground Penetrating Radar (MIMO GPR) systems working in down-looking contactless mode. The approach exploits a linear approximation of the scattering phenomenon and is based on a ray-based propagation model, which takes into account the presence of the air-soil interface. Accordingly, the Interface Reflection Point concept is extended to the case of MIMO GPR. The proposed approach performs the imaging in the 2D scalar case and applies the Truncated Singular Value Decomposition regularization scheme to perform the inversion. The effectiveness of the approach is assessed by processing synthetic and real data. The real data are referred to the lunar soil and have been collected by means of the Lunar Regolith Penetrating Radar, installed on the Chang’E-5 lander.

This paper proposes a three-dimensional imaging approach for contactless ground penetrating radar surveys. The imaging problem is formulated in the linear inverse scattering context and solved by using the Singular Values Decomposition tool. A ray-based model accounting for the electromagnetic signal propagation into an inhomogeneous medium is developed to accurately evaluate the kernel of the integral equation to be inverted. Under the proposed model, an analysis of the spatial resolution performance is carried out as a function of the geometrical and electromagnetic parameters of the scenario. To this end, theoretical concepts based on diffraction tomography and the Singular Value Decomposition of the scattering operator are exploited. Reconstruction results based on full-wave simulated data assess the feasibility of the imaging approach.

This letter deals with an imaging strategy able to manage effectively data collected on multiple lines by means of a Mini-Unmanned Aerial Vehicle (M-UAV) radar system operating in Sounder mode. The strategy, named multilines imaging approach (MIA), allows an effective 3-D pseudo-representation of the investigated volume. At the first step, MIA faces the problem of reconstructing 2-D domains (slices) by exploiting data collected on one or more lines. Then, MIA interpolates the 2-D reconstructions to provide the 3-D representation. The imaging of each slice is formulated as a linear inverse scattering problem, which is solved by means of the truncated singular value decomposition (TSVD) regularization scheme. The MIA effectiveness is assessed by processing real data collected at Archeological Park of Paestum and Velia, Paestum, Italy.