Applied Intelligence最新文献

With the networking of industrially deployed facilities in distributed environments, industrial control systems (ICS) are facing an escalating number of attacks, emphasizing the criticality of intrusion detection systems. Currently, machine learning-based intrusion detection systems have been extensively researched. However, the sensitivity of ICS data poses a challenge of scarce labeled data for these systems. Additionally, distributed ICS necessitate privacy-preserving collaborative detection. To address these challenges, some solutions combining federated learning and transfer learning have been proposed. Nonetheless, these solutions often overlook the clustering characteristics of factory equipment and the constraints posed by limited computational and communication resources. Therefore, we propose GC-FADA, a chained cross-domain collaborative intrusion detection framework, to effectively address the interplay between labeled data scarcity, privacy protection, and resource constraints in ICS intrusion detection techniques. Firstly, GC-FADA used the adversarial domain adaptation scheme to train the local model to alleviate the performance limitation of intrusion detection model caused by labeled data scarcity. Then, to reduce the communication overhead between the nodes in the factory communication network and protect client privacy, GC-FADA utilizes the geographical clustering characteristics of the factory devices and proposes a FL-based grouped chain learning structure to achieve collaborative training. Finally, GC-FADA achieves privacy protection with low computational overhead by utilizing patterns from lightweight pseudo-random generators instead of complex cryptographic primitives. Extensive experiments conducted on real industrial SCADA datasets validate the effectiveness and rationality of the proposed approach, proving that GC-FADA outperforms major domain adaptation methods in terms of accuracy while reducing computation and communication costs. In the cross-domain learning task on the two data sets, the detection accuracy of our GC-FADA reaches 88.7% and 98.29% respectively, and the detection accuracy of various network attacks is mostly more than 90%.

LiDAR and camera are two key sensors that provide mutually complementary information for 3D detection in autonomous driving. Existing multimodal detection methods often decorate the original point cloud data with camera features to complete the detection, ignoring the mutual fusion between camera features and point cloud features. In addition, ground points scanned by LiDAR in natural scenes usually interfere significantly with the detection results, and existing methods fail to address this problem effectively. We present a simple yet efficient anchor-free 3D object detection, which can better adapt to complex scenes through the adaptive fusion of multimodal features. First, we propose a fully convolutional bird’s-eye view reconstruction module to sense ground map geometry changes, for improving the interference of ground points on detection results. Second, a multimodal feature adaptive fusion module with local awareness is designed to improve the mutual fusion of camera and point cloud features. Finally, we introduce a scale-aware mini feature pyramid networks (Mini-FPN) that can directly regress 3D bounding boxes from the augmented dense feature maps, boosting the network’s ability to detect scale-varying objects, and we additionally construct a scene-adaptive single-stage 3D detector in an anchor-free manner. Extensive experiments on the KITTI and nuScenes datasets validate our method’s competitive performance.

Deep learning networks typically require vast amounts of labeled data for effective training. However, recent research has introduced a challenging task called One-Shot Object Detection, which addresses scenarios where certain classes are novel and unseen during training and represented by only a single labeled example. In this paper, we propose a novel One-Shot Object Detection model applicable to Conditional Detection without over-training on novel classes. Our approach leverages the strengths of YOLOv8 (You Only Look Once v8), a popular real-time object detector. Specifically, we incorporate a Siamese network and a matching module to enhance One-Shot Object Detection capabilities. Our proposed model, SiamYOLOv8, enables exploration of new applications without being limited by its training data. To evaluate the performance, we introduce a novel methodology for using the Retail Product Checkout (RPC) dataset “(https://github.com/MatD3mons/Conditional-Detection-datasets/tree/main/RPC)”, and extend our evaluation using the Grozi-3.2k dataset “(https://github.com/MatD3mons/Conditional-Detection-datasets/tree/main/GROZI-3.2k)”. In such contexts, new products often lack sufficient data for continuous Deep Learning methods, making individual case identification difficult. Our model outperforms SOTA models, achieving a significant performance improvement of 20.33% increase in Average Precision (+12.41 AP) on the Grozi-3.2k dataset and 25.68% increase (+17.37 AP) on the RPC dataset.

Class Incremental Semantic Segmentation (CISS) aims at segmenting the incremental new classes without losing the ability on old classes. Currently, some CISS methods based on feature knowledge distillation suffer from the stability-plasticity dilemma, i.e., excessive knowledge distillation may impede models from learning new classes. Besides, distilling without emphasis fails to preserve old knowledge effectively. To address these issues, a more fine-grained and focused approach to knowledge transfer, named dual attention-guided distillation (DAGD), is proposed for the CISS task. This approach not only ensures that the inherited knowledge is distilled in a targeted manner but also allows the model to adapt and learn new knowledge more efficiently. DAGD model contains a channel attention-guided distillation module and a spatial attention-guided distillation module. The former distills channel-wise attention maps to improve the knowledge transfer of essential channels while accommodating new knowledge learning. The latter encodes a weight coefficient map to highlight important regions in the spatial dimension, which further decouples old knowledge retention and new knowledge entry. Furthermore, a dynamic temperature strategy is introduced to facilitate logit knowledge distillation, specifically sharpening the predictive distribution produced by the output of the old model, thus achieving more accurate knowledge transfer. Extensive experimental results on Pascal VOC 2012 and ADE20K datasets demonstrate that our method achieves competitive results.

In recent years, Transformers have revolutionized the management of Natural Language Processing tasks, and Vision Transformers (ViTs) promise to do the same for Computer Vision ones. However, the adoption of ViTs is hampered by their computational cost. Indeed, given an image divided into patches, it is necessary to compute for each layer the attention of each patch with respect to all the others. Researchers have proposed many solutions to reduce the computational cost of attention layers by adopting techniques such as quantization, knowledge distillation and manipulation of input images. In this paper, we aim to contribute to the solution of this problem. In particular, we propose a new framework, called AgentViT, which uses Reinforcement Learning to train an agent that selects the most important patches to improve the learning of a ViT. The goal of AgentViT is to reduce the number of patches processed by a ViT, and thus its computational load, while still maintaining competitive performance. We tested AgentViT on CIFAR10, FashionMNIST, and Imagenette(^+) (which is a subset of ImageNet) in the image classification task and obtained promising performance when compared to baseline ViTs and other related approaches available in the literature.

Globally, cyberattacks are growing and mutating each month. Intelligent Intrusion Network Detection Systems are developed to analyze and detect anomalous traffic to face these threats. A way to address this is by using network flows, an aggregated version of communications between devices. Network Flow datasets are used to train Artificial Intelligence (AI) models to classify specific attacks. Training these models requires threat samples usually generated synthetically in labs as capturing them on operational network is a challenging task. As threats are fast-evolving, new network flows are continuously developed and shared. However, using old datasets is still a popular procedure when testing models, hindering a more comprehensive characterization of the advantages and opportunities of recent solutions on new attacks. Moreover, a standardized benchmark is missing rendering a poor comparison between the models produced by algorithms. To address these gaps, we present a benchmark with fourteen recent and preprocessed datasets and study seven categories of algorithms for Network Intrusion Detection based on Network Flows. We provide a centralized source of pre-processed datasets to researchers for easy download. All dataset are also provided with a train, validation and test split to allow a straightforward and fair comparison between existing and new solutions. We selected open state-of-the-art publicly available algorithms, representatives of diverse approaches. We carried out an experimental comparison using the Macro F1 score of these algorithms. Our results highlight each model operation on dataset scenarios and provide guidance on competitive solutions. Finally, we discuss the main characteristics of the models and benchmarks, focusing on practical implications and recommendations for practitioners and researchers.

Understanding the causal relationships between data variables can provide crucial insights into the construction of tabular datasets. Most existing causality learning methods typically focus on applying a single identifiable causal model, such as the Additive Noise Model (ANM) or the Linear non-Gaussian Acyclic Model (LiNGAM), to discover the dependencies exhibited in observational data. We improve on this approach by introducing a novel dual-step framework capable of performing both causal structure learning and tabular data synthesis under multiple causal model assumptions. Our approach uses Directed Acyclic Graphs (DAG) to represent causal relationships among data variables. By applying various functional causal models including ANM, LiNGAM and the Post-Nonlinear model (PNL), we implicitly learn the contents of DAG to simulate the generative process of observational data, effectively replicating the real data distribution. This is supported by a theoretical analysis to explain the multiple loss terms comprising the objective function of the framework. Experimental results demonstrate that DAGAF outperforms many existing methods in structure learning, achieving significantly lower Structural Hamming Distance (SHD) scores across both real-world and benchmark datasets (Sachs: 47%, Child: 11%, Hailfinder: 5%, Pathfinder: 7% improvement compared to state-of-the-art), while being able to produce diverse, high-quality samples.

Social networking platforms have been generating a massive amount of data in real-time that can be analysed and used to support government and relief organizations in preparing quick and effective action plans for disaster response. Effective disaster response requires a broad understanding of disaster situations, such as the emergency necessities of the people, their sentiments towards emergency needs, and the geographical distribution of their requirements and opinions. However, in literature, many studies exist that estimate the emotions and sentiments of the people during a disaster; they are inept in identifying and mapping the public sentiments toward emergency needs. This paper proposes a framework called SentimentMapper. This framework quickly maps the sentiments of people toward emergency needs using social media data to plan for effective disaster response. In order to perform an automatic analysis of sentiments using Twitter (re-branded to X since July 2023) data, we introduce a BERT Convolutional Neural Network (BCNN). BCNN performs the sentiment analysis of the collected data from the disaster-affected people regarding essential needs like food, shelter, medical emergency, and rescue during different disasters. Next, we present a tweet-text independent approach to detect the location of the tweets posted on Twitter and discover the impacts in different areas due to any disaster event. Furthermore, we also study the variations in public attitudes about the essential needs during identical or different disasters. As a case study, the proposed framework has been used on the dataset collected from Twitter during the Assam flood 2021 in India and validated with the corresponding survey reports published by the government agency. The detailed results of the analytics in the proposed framework and its validation with the case study data confirm that it is capable of providing credible situational information quickly required for the disaster responses.

Model reparameterization is a widely accepted technique for improving inference speed without compromising performance. However, current Post-training Quantization (PTQ) methods often lead to significant accuracy degradation when applied to reparameterized models. This is primarily caused by channel-specific and sample-specific outliers, which appear only at specific samples and channels and impact on the selection of quantization parameters. To address this issue, we propose RepAPQ, a novel framework that preserves the accuracy of quantized reparameterization models. Different from previous frameworks using Mean Squared Error (MSE) as a measurement, we utilize Mean Absolute Error (MAE) to mitigate the influence of outliers on quantization parameters. Our framework consists of two core components: Quantization Protecting Reparameterization and Across-block Calibration. For effective calibration, Quantization Protecting Reparameterization combines multiple branches into a single convolution with an affine layer. During training, the affine layer accelerates convergence and amplifies the output of the convolution to better accommodate samples with outliers. Additionally, Across-block Calibration leverages the measurement of stage output as supervision to address the gradient problem introduced by MAE and enhance the interlayer correlation with quantization parameters. Comprehensive experiments demonstrate the effectiveness of RepAPQ across various models and tasks. Our framework outperforms previous methods by approximately 1% for 8-bit PTQ and 2% for 6-bit PTQ, showcasing its superior performance. The code is available at https://github.com/ilur98/DLMC-QUANT.

Stock price prediction remains a critical challenge in financial research due to its potential to inform strategic decision-making. Existing approaches predominantly focus on two key tasks: (1) regression, which forecasts future stock prices, and (2) classification, which identifies trading signals such as buy, sell, or hold. However, the inherent limitations of financial data hinder effective model training, often leading to suboptimal performance. To mitigate this issue, prior studies have expanded datasets by aggregating historical data from multiple companies. This strategy, however, fails to account for the unique characteristics and interdependencies among individual stocks, thereby reducing predictive accuracy. To address these limitations, we propose a novel BiLSTM-GAT-AM model that integrates bidirectional long short-term memory (BiLSTM) networks with graph attention networks (GAT) and an attention mechanism (AM). Unlike conventional graph-based models that define edges based solely on technical or fundamental relationships, our approach employs a dual-graph structure: one graph captures technical similarities, while the other encodes fundamental industry relationships. These two representations are aligned through an attention mechanism, enabling the model to exploit both technical and fundamental insights for enhanced stock market predictions. We conduct extensive experiments, including ablation studies and comparative evaluations against baseline models. The results demonstrate that our model achieves superior predictive performance. Furthermore, leveraging the model’s forecasts, we construct an optimized portfolio and conduct backtesting on the test dataset. Empirical results indicate that our portfolio consistently outperforms both baseline models and the S&P 500 index, highlighting the effectiveness of our approach in stock market prediction and portfolio optimization.