The widespread integration of Artificial Intelligence (AI) in sixth-generation Internet of Things (6G-IoT) applications, introduces significant challenges for ensuring the trustworthy and dependability of AI models. The "black-box" characteristic of numerous Deep Neural Networks (DNNs) creates a notable obstacle for confirming their safety in intricate, ever-changing environments. Consequently, there is a need for extensive testing, requiring the gathering and labeling of a large number of test cases, a process that is both time-intensive and resource-consuming. While previous studies have adapted neuron coverage criteria for steering test case generation in DNNs. Yet, these criteria are white-box measures requiring access to model states and presenting their practical limitations. Conversely, black-box metrics, which focus on outputs, present a more feasible approach. Among these, black-box diversity metrics evaluate model robustness by generating diverse test cases, eliminating the need for internal model details. This paper presents a test case generation framework centered on diversity, known as BDTest. BDTest enhances test adequacy through five stages: (1) Mapping feature vectors extracted from an initial set of seed images onto a low-dimensional manifold utilizing UMAP; (2) Detecting sparse regions using DBSCAN; (3) Sampling key points from these regions via Latin Hypercube Sampling; (4) Reconstructing latent features and generating new images through ICA and GAN inversion; and (5) Measuring the diversity of the generated set using metrics such as the Log-Determinant. Experiments demonstrate that BDTest significantly improves test set diversity and error detection performance, achieving error rates of 59.36%, 59.76%, and 67.03% on VGG19, DenseNet121, and MobileNetV2, respectively, outperforming DeepXplore by an average of 12.43% and DLFuzz by 9.95% across all tested models. When retrained with the generated test cases, the model demonstrated improved accuracy on the original test set, alongside a significant enhancement in accuracy on the natural adversarial test set.

Traffic flow prediction plays a crucial role in the construction of smart cities. Although numerous models already exist for traffic flow prediction, they neither extracted the spatio-temporal features at different scales nor precisely considered the deviation between the traffic signals collected by the sensors and the trend signals at different scales. This leads to their inaccurate prediction results. To address the aforementioned issues, this paper proposes a multi-layer structure to extract spatio-temporal features at different scales and designs an information decomposition module to separate abnormal signals in traffic data. Furthermore, based on the above structure and modules, this paper constructs a new traffic flow prediction modelMulti-scaleAnomalyDecompositionGraphNeuralNetwork (MADGNN) for feature extraction and information decomposition at different scales. Firstly, the model encodes the input data to fully capture spatio-temporal dependencies. Then, this paper extracts spatio-temporal features at multiple scales based on a multi-layer structure containing multiple GRUs and subtracts the learned abnormal information from the input signal to achieve abnormal signal decomposition. Finally, we use multiple spatio-temporal hidden states for further information extraction and traffic prediction. The final prediction result of the model is obtained by adding up the prediction outputs of each layer. The experimental results show that, compared with DDGCRN, on the Railway datasets, the MAE and RMSE metrics are improved by an average of 3.92% and 2.30% respectively, and on the public dataset PEMSD8, the MAE and MAPE metrics are improved by 2.92% and 2.23% respectively.

Federated learning (FL) is a novel paradigm for distribute edge intelligence for the Internet-of-Vehicles (IoV) application, which can enable superior performance in model training without the need to share local data. However, in the actual architecture of FL, the existence of nonindependent and identically distributed (non-IID) data at the edge device, along with the involvement of randomly participating distributed nodes, can result in model bias and a subsequent decrease in overall performance. To solve this problem, a new federated self-supervised learning method based on prototypes clustering contrastive learning (FedPCC) is proposed, which can effectively addresses the issue of asynchronous edge training and global model bias by introducing an unsupervised prototypes layer. The prototypes layer maps edge features to a global space and performs clustering, facilitating the new aggregation method of global prototypes on the server. Then, models from other components are aggregated based on data weight. Besides that, during the parameter deployment phase, we replace the prototype layer to acquire global knowledge, while employing momentum updates to preserve the local knowledge of the other components. Finally, to assess the efficacy of our proposed approach, we carried out comprehensive experiments across the various data sets. The findings show that our method gains state-of-the-art performance, which also validates its effectiveness.

Spatiotemporal signal forecasting is vital for promoting intelligence management in Internet of Everything applications. Benefiting from the powerful representation ability of deep learning, DNN based methods have shown promising performance in spatiotemporal forecasting. However existing methods perform suboptimally in few-shot distribution shift scenarios. As the representation ability results in stable fitting, the model struggles to adapt to discrepancy in the data domain. And it is challenging to modify the mapping relationship between the data and the latent space of representation with few-shot distributionally shifted target data. In this case, we propose a distribution centric prompt based transfer learning framework, which transforms the data distribution information into model-interpretable prompt embeddings confused with spatiotemporal sematic information in Intermediary Bridging Space which serves as a mediator between the data domain and the latent space. Thus the downstream model can learn a target distribution aligned representation regulated by the mediator. Though experiments on real-world datasets, we verify the effectiveness and extensibility of the proposed method.

The upcoming 6G technology, with its high speed and low latency, is poised to become a foundational technology for intelligent transportation systems. To handle the massive data generated by connected vehicles in 6G environments, federated learning methods are essential. However, traditional centralized federated learning approaches still face challenges related to data and device heterogeneity, which significantly affects training efficiency. To address these challenges, we propose FedCPC, a context-based adaptive pruning clustered federated learning method. Based on the positive correlation between similar data distributions and model representations, we use centralized kernel alignment (CKA) to group clients with similar data distributions, thus reducing the impact of data heterogeneity. Furthermore, we introduce a context-aware random forest multi-armed bandit method to determine appropriate pruning rates based on device capabilities and historical performance which addresses device heterogeneity concerns. Experimental results on open-source datasets demonstrate that FedCPC outperforms traditional FL methods in both learning efficiency and communication effectiveness.

Recently, deep learning systems have been widely utilized in various fields, prompting increased attention to their security. Fuzz testing is a crucial automated testing method; however, traditional approaches are not directly applicable to the testing of deep neural networks (DNNs). In light of this challenge, this study proposes a DNN fuzz testing method based on gradient weighted class activation graphs. By integrating model visualization interpretation technology and Grad-CAM technology, only significant areas are disrupted to rapidly generate test cases capable of inducing DNN errors. Additionally, high-quality initial seeds are selected based on the heat map to assess the degree of image feature distinctiveness. Adversarial perturbations are then exclusively applied to areas with high heat values in order to enhance the authenticity of the generated images. Experimental results demonstrate that this approach effectively enhances model robustness and accuracy, produces high-quality test cases, and significantly contributes to model repair efforts.

Semantic segmentation constitutes an essential component of various visual tasks and is widely utilized in numerous applications. However, its implementation in real-world settings is frequently challenged by limitations pertaining to resources and latency. The early exit mechanism addresses these challenges by attaching Internal Segmentation Heads (ISHs) to intermediate layers of the network, facilitating a gradual segmentation of the entire image from less complex to more intricate regions. This methodology mitigates the aggregate computational burden by obviating the necessity to execute the entire network across all portions of the image. However, in antecedent methodologies, ISHs function in isolation from each other. This segregation engenders complications in which the latter ISHs, designated to manage more intricate and detailed regions, are deprived of access to high-resolution information. Furthermore, the subsequent ISHs are unable to fully capitalize on the computations executed by preceding ISHs, thereby resulting in inefficient utilization of computational resources. To ameliorate this challenge, we introduce the Cascade-Exit Semantic Segmentation Network (CESS). This innovative architecture propagates high-resolution features into the decoders of subsequent ISHs and establishes connections between the predictors of antecedent ISHs and those of the subsequent ones. Comprehensive empirical evaluations performed on the Cityscapes and CamVid datasets substantiate that our proposed methodology exhibits a superior balance between accuracy and inference time when juxtaposed with recently early exit techniques.

Deep neural networks (DNN) have extensive applications in image processing, medical diagnosis, autonomous driving, and various other domains. Previous research has found limitations in the error-awareness capacities of deep neural networks, especially when deployed in safety-critical systems. Several neuron coverage criteria have been suggested to direct the adequate testing and quality assessment of deep neural networks, drawing inspiration from conventional software quality assessment methodologies. Nevertheless, recent research has raised doubts about the validity of these criteria in guiding quality assessment of deep neural networks and measuring test adequacy. In this paper, we use ImageNet-A as an additional test dataset and employ three diversity metrics to assess the diversity of the hybrid datasets comprising both ImageNet and ImageNet-A. Subsequently, we employ the hybrid dataset in conjunction with eight DNNs, to investigate the statistical correlation between the three diversity metrics and the DNNs’ accuracy in predicting correct outcomes. Furthermore, we conduct a comparative analysis between the diversity metrics and DNN accuracy, as well as between coverage criteria and DNN accuracy. The classification task used in the experiment is also widely used in various sensors for ubiquitous intelligence. The results of our experiment indicate that test set diversity metrics serve as superior indicators of test adequacy compared to coverage criteria, thereby enabling more effective guidance for testing DNNs.

The emerging field of Artificial Intelligence of Things (AIoT) is propelled by the remarkable success of deep learning and hardware evolution, which has a significant impact on our daily lives. However, because of their notorious computing resource intensity, the widespread deployment of AIoT devices requires substantial electricity consumption as support, inevitably escalating energy consumption, and ultimately leads to a significant carbon emissions. Existing research on neural network compression and acceleration struggles to achieve energy-efficient inference on resource-constrained AIoT devices. To address this issue, we propose a software-hardware co-design approach that integrates advanced neural network optimization techniques with hardware power management capabilities to enable energy-efficient inference and ultimately achieve sustainable AIoT. We introduce a lightweight split and refinement block that adaptively reduces redundant computation in both channel and spatial dimensions. Several early exit (EE) branches are added to the backbone, which are controlled by a policy-based EE predictor. With the predicted EE index, a curve-fitting-based frequency scaling algorithm is presented to calculate the optimal frequency that minimizes energy overhead while maintaining latency constraints. Extensive experiments on CIFAR and CINIC classification tasks validate that our proposed method consistently reduces energy consumption for neural network inference while outperforming other competitive methods.

Quantum circuit mapping (QCM) is a crucial preprocessing step for executing a logical circuit (LC) on noisy intermediate-scale quantum (NISQ) devices. Balancing the introduction of extra gates and the efficiency of preprocessing poses a significant challenge for the mapping process. To address this challenge, we propose the mutual-influence-aware (MIA) heuristic method by integrating an initial mapping search framework, an initial mapping generator, and a heuristic circuit mapper. Initially, the framework utilizes the generator to obtain a favorable starting point for the initial mapping search. With this starting point, the search process can efficiently discover a promising initial mapping within a few bidirectional iterations. The circuit mapper considers mutual influences of SWAP gates and is invoked once per iteration. Ultimately, the best result from all iterations is considered the QCM outcome. The experimental results on extensive benchmark circuits demonstrate that, compared to the iterated local search (ILS) method, which represents the current state-of-the-art, our MIA method introduces a similar number of extra gates while achieving nearly 95 times faster execution.