Gradient smoothing is an efficient approach to reducing noise in gradient-based model explanation methods. SmoothGrad adds Gaussian noise to mitigate much of this noise. However, the crucial hyperparameter in this method, the variance $\sigma$ of the Gaussian noise, is often set manually or determined using a heuristic approach. This results in the smoothed gradients containing extra noise introduced by the smoothing process. In this paper, we aim to analyze the noise and its connection to the out-of-range sampling in the smoothing process of SmoothGrad. Based on this insight, we propose AdaptGrad, an adaptive gradient smoothing method that controls out-of-range sampling to minimize noise. Comprehensive experiments, both qualitative and quantitative, demonstrate that AdaptGrad could effectively reduce almost all the noise in vanilla gradients compared to baseline methods. AdaptGrad is simple and universal, making it a practical solution to enhance gradient-based interpretability methods to achieve clearer visualization.

Time series anomaly detection is one kind of critical time series analytical tasks, which is widely applied to various real-world applications. Recently, the diffusion models have shown promising performance on time series imputation for anomaly detection. And the high computing requirements of the diffusion models naturally extend their computation paradigm from the centralized computing to the cloud computing, and then further to JointCloud computing which allows the diffusion models to be deployed across multiple clouds. However, the JointCloud computing paradigm faces the challenge of task offloading over multiple nodes across multiple clouds. Unfortunately, most of existing task offloading methods overlook the dynamic nature of network conditions among clouds. To address this issue, we propose a time series anomaly detection approach named AdaDiffAD by adaptively segmenting diffusion models in JointCloud environment with dynamic network conditions. Specifically, we design a task offloading strategy by segmenting the denoising process of the diffusion model onto both the edge clouds and central clouds, and utilizing the edge cloud results directly for anomaly detection when the network condition is not ideal. By conducting comprehensive experiments on seven datasets, the experimental results demonstrate that our proposed AdaDiffAD always achieves lower time consumption while maintaining competitive anomaly detection performance compared to the state-of-the art without adaptive task offloading strategy.

Anomaly detection in multivariate time series has emerged as a critical challenge in the time series research community with significant application potentials in various scenarios, ranging from fault diagnosis to system state estimation in Industrial Control Systems (ICSs). Meanwhile, the demand for high availability and extensibility of ICSs necessitates their deployment in the JointCloud environment. Therefore, the performance of the multivariate time series anomaly detection model is expected to be enhanced in the JointCloud environment when encountering dynamic network conditions among multiple clouds. Impressed by the effectiveness of diffusion models in anomaly detection, we have chosen diffusion models for empowering our anomaly detection model. Specifically, we propose Progressively Distilled Diffusion Anomaly Detection model (ProDiffAD) in the JointCloud environment to seek for the balance between effectiveness and efficiency. Moreover, our proposed model is capable of being adaptive with the dynamic network conditions in the JointCloud environment by modeling the intercloud network conditions. To validate the effectiveness and efficiency of our model, comprehensive experiments are conducted on two real and five synthetic datasets. The experimental results demonstrate that our proposed model achieves more accurate and faster multivariate time series anomaly detection in the JointCloud environment under dynamic network conditions compared to state-of-the-art models.

Abstract In the medical image processing domain, deep learning methodologies have outstanding performance for disease classification using digital images such as X‐rays, magnetic resonance imaging (MRI), and computerized tomography (CT). However, accurate diagnosis of disease by medical personnel can be challenging in certain cases, such as the complexity of interpretation and non‐availability of expert personnel, difficulty at pixel‐level analysis, etc. Computer‐aided diagnostic (CAD) systems with proper training have shown the potential to enhance diagnostic accuracy and efficiency. With the exponential growth of medical data, CAD systems can analyze and extract valuable information by assisting medical personnel during the disease diagnostic process. To overcome these challenges, this research introduces CX‐RaysNet, a novel deep‐learning framework designed for the automatic identification of various lung disease classes in digital chest X‐ray images. The core novelty of the CX‐RaysNet framework lies in the integration of both convolutional and group convolutional layers, along with the usage of small filter sizes and the incorporation of dropout regularization. This phenomenon helps us optimize the model's ability to distinguish minute features that reveal different lung diseases. Additionally, data augmentation techniques are implemented to augment the training and testing datasets, which enhances the model's robustness and generalizability. The performance evaluation of CX‐RaysNet reveals promising results, with the proposed model achieving a multi‐class classification accuracy of 97.25%. Particularly, this study represents the first attempt to optimize a model specifically for low‐power embedded devices, aiming to improve the accuracy of disease detection while minimizing computational resources.

Deepfake threatens the authenticity of the information in artificial intelligence Internet of Things (IoT) systems. Recently, several deepfake detection methods have been proposed in academia and industry for securing the authenticity of visual information in the face of artificial intelligence advances. Frame-level detection methods, a widely employed security method against deepfake, have a small model size and offer real-time responsiveness, despite basing their classification decision only on the information contained within the frame they are evaluating. We propose a new lightweight frame-level detection technique based on Bayesian inference weighting (BIW) to improve the robustness of existing frame-level detection models. Our proposed BIW technique employs the Naive Bayesian algorithm to estimate the reliability of any candidate model’s detection results. Comprehensive experiments were conducted on the attacked data sets by four designed video interference approaches and edge computing platform, showing that BIW enhances the robustness of all the baselines and improves their detection accuracy with a real-time response.

How to satisfy safety constraints almost surely (or with probability one) is becoming an emerging research issue for safe reinforcement learning (RL) algorithms in safety-critical domains. For instance, self-driving cars are expected to ensure that the driving strategy they adopt will never do harm to pedestrians and themselves. However, existing safe RL algorithms suffer from either risky and unstable constraint satisfaction or slow convergence. To tackle these two issues, we propose Conservatively Constrained Policy Optimization (CCPO) using state augmentation. CCPO designs a simple yet effective penalized reward function by introducing safety states and adaptive penalty factors under Safety Augmented MDP framework. Specifically, a novel Safety Promotion Function (SPF) is proposed to make the agent being more concentrated on constraint satisfaction with faster convergence by reshaping a more conservative constrained optimization objective. Moreover, we theoretically prove the convergence of CCPO. To validate both the effectiveness and efficiency of CCPO, comprehensive experiments are conducted in both single-constraint and more challenging multi-constraint environments. The experimental results demonstrate that the safe RL algorithms augmented by CCPO satisfy the predefined safety constraints almost surely and gain almost equivalent cumulative reward with faster convergence.

Enset (Ensete ventricosum (Welw.) Cheesman) is an indigenous Ethiopian crop that sustains the livelihood of more than 20 million people. Enset is used as a crop for human food security, animal feed, and a source of fiber for farmers. Since the past few decades, the production of enset has been severely curbed by bacterial wilt disease. The early detection and prevention of this disease are crucial for enhancing production. Deep learning in plant disease management is becoming an effective way to improve agro-productivity. However, standard convolutional neural network (CNN) models require a large number of parameters and higher computational costs. Efficient CNN models allow users to benefit without having to submit their data to a server for analysis, which is especially useful in parts of the world where internet access is fragile or even inaccessible. In this paper, we proposed a MobileNetV3-Small model which we jointly trained newly added classifier layers with selected final layers of the base model. The proposed model achieved 99.93% of accuracy rate on a test set. The model is trained using a dataset that includes 99 popular enset clones we collected from five different locations in Ethiopia with varying altitudes, climates, and weather conditions.

Emotion cause pair extraction (ECPE), as one of the derived subtasks of emotion cause analysis (ECA), shares rich inter-related features with emotion extraction (EE) and cause extraction (CE). Therefore EE and CE are frequently utilized as auxiliary tasks for better feature learning, modeled via multi-task learning (MTL) framework by prior works to achieve state-of-the-art (SoTA) ECPE results. However, existing MTL-based methods either fail to simultaneously model the specific features and the interactive feature in between, or suffer from the inconsistency of label prediction. In this work, we consider addressing the above challenges for improving ECPE by performing two alignment mechanisms with a novel Aˆ2Net model. We first propose a feature-task alignment to explicitly model the specific emotion-&cause-specific features and the shared interactive feature. Besides, an inter-task alignment is implemented, in which the label distance between the ECPE and the combinations of EE&CE are learned to be narrowed for better label consistency. Evaluations of benchmarks show that our methods outperform current best-performing systems on all ECA subtasks. Further analysis proves the importance of our proposed alignment mechanisms for the task.

In the field of data mining, Time Series Classification (TSC) has attracted lots of interest from researchers due to its wide range of applications. Recently, deep learning models have shown promising performance on TSC by automatically extracting discriminative features from original time series. However, deep learning models are very time-consuming for training and lack interpretability. In the contrast, shapelets-based algorithms are easy to be implemented with interpretability preserved. In this paper, we proposed a salience-guided shapelets selection approach named Points2Shapelets, which inherits the advantages of both deep learning-based models and shapelets-based algorithms for TSC. Specifically, an interpretable and transparent shapelets selection approach is designed by leveraging the salience analysis of the pre-trained model. By conducting comprehensive experiments on 20 public UCR time series datasets, the experimental results demonstrate that our proposed approach Points2Shapelets achieves competitive performance shown by deep learning-based models. Meanwhile, explanations about how the salience analysis guides the decision-making of TSC models are offered in a visually understandable manner.

Temporal knowledge graph (TKG) has received great attention in recent years. However, the TKG is not always complete due to the missing of important facts, which has seriously hindered its wide application. Inferring missing facts in TKG is a critical and challenging task due to its highly dynamic nature. Most of the existing methods mainly focus on modeling the structural features and temporal dependencies of TKG to solve the temporal knowledge graph completion problem (TKGC). However, those methods only operate at a single timescale without considering the latent time variability of TKG and thus limit the performance of TKGC solutions. Therefore, we propose a novel method named MtGCN (Multi-timescale history modeling framework based on Graph Convolutional Networks) for completing TKG by self-adaptively modeling the multi-timescale history of the incomplete TKG. Firstly, MtGCN uses a structural encoder with a graph convolutional network to mine the latent semantic information and structural features of the TKG. Secondly, MtGCN uses GRU-based temporal encoder to learn the historical information at various timescales of the TKG. Finally, it generates effective entity and relation representations to infer the missing facts for the originally incomplete TKG. By conducting comprehensive experiments on 5 public datasets, the experimental results show that our proposed method MtGCN significantly outperforms the baselines by achieving the highest MRR and HITS@1,3,10.

Federated learning, as a typical distributed learning paradigm, shows great potential in Industrial Internet of Things, Smart Home, Smart City, etc. It enables collaborative learning without data leaving local users. Despite the huge benefits, it still faces the risk of privacy breaches and a single point of failure for aggregation server. Adversaries can use intermediate models to infer user privacy, or even return incorrect global model by manipulating the aggregation server. To address these issues, several federated learning solutions focusing on privacy-preserving and security have been proposed. However, theses solutions still faces challenges in resource-limited scenarios. In this paper, we propose G-VCFL, a grouped verifiable chained privacy-preserving federated learning scheme. Specifically, we first use the grouped chain learning mechanism to guarantee the privacy of users, and then propose a verifiable secure aggregation protocol to guarantee the verifiability of the global model. G-VCFL does not require any complex cryptographic primitives and does not introduce noise, but enables verifiable privacy-preserving federated learning by utilizing lightweight pseudorandom generators. We conduct extensive experiments on real-world datasets by comparing G-VCFL with other state-of-the-art approaches. The experimental results and functional evaluation indicate that G-VCFL is efficient in the six experimental cases and satisfies all the intended design goals.

Probing techniques are widely used to identify faulty nodes in networks. Existing probe-based solutions for SDN fault localizationcan focus on two ways: per-rule and per-path. Both promote some certain switches to reporters by installing on them report rules. To avoid hindering other test packets, such report rules must vary between tests or be deleted before a next test, thus incurring excessive consumption on either TCAM resources of switches or bandwidth reserved for control messages. In this paper we present Voyager, a hybrid fault localization solution for SDN that fully combines the advantages of per-rule and per-path tests. Voyager significantly reduces the number of report rules and allows them to reside and function in switches persistently. With only one well-designed report rule for each switch installed, Voyager pinpoints faulty switches easily and tightly by sending test packets straight. Tests in Voyager are parallelizable and report rules are non-invasive. The performance evaluation on realistic datasets shows that Voyager is 24.0% to 92.3% faster than existing solutions.

Missing values appearing in multivariate time series often prevent further and in-depth analysis in real-world applications. To handle those missing values, advanced multivariate time series imputation methods are expected to (1) consider bi-directional temporal correlations, (2) model cross-variable correlations, and (3) approximate original data's distribution. However, most of existing approaches are not able to meet all the three above-mentioned requirements. Drawing on advances in machine learning, we propose BRNN-GAN, a generative adversarial network with bi-directional RNN cells. The BRNN cell is designed to model bi-directional temporal and cross-variable correlations, and the GAN architecture is employed to learn original data's distribution. By conducting comprehensive experiments on two public datasets, the experimental results show that our proposed BRNN-GAN outperforms all the baselines in terms of achieving the lowest Mean Absolute Error (MAE).

Time series classification (TSC) is widely used in various real-world applications such as human activity recognition, smart city governance, etc. Unfortunately, due to different reasons, only part of time series could be collected which may obviously degrade the performance of time series classifiers. To alleviate this problem, time series augmentation aims to generate synthetic time series by learning useful features from collected time series. As the popular generative model, generative adversarial networks (GAN) is regarded as a promising model for time series augmentation. However, applying GAN to the time series data suffers from a challenge in which the generated instances hold low quality but the model has gotten saturation. In this paper, for time series augmentation, we proposed TSA-GAN which is a robust GAN model with a self-adaptive recovering strategy to solve this problem. On 85 datasets of the UCR 2015 archive, our proposed TSA-GAN helps time series classifiers achieve performance improvements ranging from 8.3% to 12.5%, which is far better than the baseline.

Due to the dynamic characteristics of news and user preferences, personalized recommendation is a challenging problem. Traditional recommendation methods simply focus on current reward, which just recommend items to maximize the number of current clicks. And this may reduce users' interest in similar items. Although the news recommendation framework based on deep reinforcement learning preciously proposed (i.e, DRL, based on deep Q-learning) has the advantages of focusing on future total rewards and dynamic interactive recommendation, it has two issues. First, its exploration method is slow to converge, which may bring new users a bad experience. Second, it is hard to train on off-line data set because the reward is difficult to be determined. In order to address the aforementioned issues, we propose a framework named DEN-DQL for news recommendation based on deep Q-learning with double exploration networks. Also, we develop a new method to calculate rewards and use an off-line data set to simulate the online news clicking environment to train DEN-DQL. Then, the well trained DEN-DQL is tested in the online environment of the same data set, which demonstrates at least 10% improvement of the proposed DEN-DQL.

In recent years, Convolutional Neural Networks (CNN s) have been widely applied in various applications due to its powerful learning capability. However, its lack of explainability hinders its further usage in tasks requiring high reliability. Therefore, interpretability technique is the key to the application and deployment of CNN models. As a typical interpretability technique for CNN, Class Activation Map (CAM) utilizing the gradient based weights and activation map is widely applied to traditional CNN models for offering visual interpretability. However, the activation map adopted by CAM cannot loyally quantify the relevance between input samples and activation values. Hence, in this paper, we propose a new interpretability approach called Salience-CAM employing salience scores to accurately measure the relevance between input samples and activation values. To evaluate the effectiveness of Salience-CAM, comprehensive experiments are conducted on 6 selected time series datasets. By leveraging an evaluation algorithm proposed in this paper, the experimental results show that our proposed Salience-CAM outperforms the baseline by discovering more discriminative features.

As a typical task of time series mining, Time Series Classification (TSC) has attracted lots of attention from both researchers and domain experts due to its broad applications. To get rid of costly hand-crafting feature engineering process, deep learning techniques are applied for automatic feature extraction, which shows competitive or even better performance compared with state-of-the-art TSC solutions. However, on time series datasets presenting complex patterns, neither 1-Nearest-Neighbour classifier nor deep learning models are capable of achieving satisfactory classification accuracy which motivates us to explore new time series representations to help classifiers further improve the classification accuracy. In this paper, by building the binary distribution tree, an approach to time series classification based on deep learning models using new representations is proposed. By conducting comprehensive experiments over 6 most challenging time series datasets and comparing experimental results of the same classifier using the proposed representation or not, the potential of the proposed approach to enhancing time series classification accuracy is validated with a bunch of helpful findings.