International Journal of Machine Learning and Cybernetics最新文献

Machine learning as a service (MLaaS) has become a widely adopted approach, allowing customers to access even the most complex machine learning models through a pay-per-query model. Black-box distribution has been widely used to keep models secret in MLaaS. However, even with black-box distribution alleviating certain risks, the functionality of a model can still be compromised when customers gain access to their model’s predictions. To protect the intellectual property of model owners, we propose an effective defense method against model stealing attacks with the localized stochastic sensitivity (LSS), namely LSSMSD. First, suspicious queries are detected by employing an out-of-distribution (OOD) detector. Addressing a critical issue with many existing defense methods that overly rely on OOD detection results, thus affecting the model’s fidelity, we innovatively introduce LSS to solve this problem. By calculating the LSS of suspicious queries, we can selectively output misleading predictions for queries with high LSS using an misinformation mechanism. Extensive experiments demonstrate that LSSMSD offers robust protections for victim models against black-box proxy attacks such as Jacobian-based dataset augmentation and Knockoff Nets. It significantly reduces accuracies of attackers’ substitute models (up to 77.94%) while yields minimal impact to benign user accuracies (average (-2.72%)), thereby maintaining the fidelity of the victim model.

Circular RNAs (circRNAs) are a special class of endogenous non-coding RNA molecules with a closed circular structure. Numerous studies have demonstrated that exploring the association between circRNAs and diseases is beneficial in revealing the pathogenesis of diseases. However, traditional biological experimental methods are time-consuming. Although some methods have explored the circRNA associated with diseases from different perspectives, how to effectively integrate the multi-perspective data of circRNAs has not been well studied, and the feature aggregation between heterogeneous nodes has not been fully considered. Based on these considerations, a novel computational framework, called CHNSCDA, is proposed to efficiently forecast unknown circRNA-disease associations(CDAs). Specifically, we calculate the sequence similarity and functional similarity for circRNAs, as well as the semantic similarity for diseases. Then the similarities of circRNAs and diseases are combined with Gaussian interaction profile kernels (GIPs) similarity, respectively. These similarities are fused by taking the maximum values. Moreover, circRNA-circRNA associations and disease-disease associations with strong correlations are selectively combined to construct a heterogeneous network. Subsequently, we predict the potential CDAs based on the multi-head dynamic attention mechanism and multi-layer convolutional neural network. The experimental results show that CHNSCDA outperforms the other four state-of-the-art methods and achieves an area under the ROC curve of 0.9803 in 5-fold cross validation (5-fold CV). In addition, extensive ablation comparison experiments were conducted to confirm the validity of different similarity feature aggregation methods, feature aggregation methods, and dynamic attention. Case studies further demonstrate the outstanding performance of CHNSCDA in predicting potential CDAs.

Scanned 3D point cloud data is typically noisy and incomplete. Existing point cloud completion methods tend to learn a mapping of available parts to the complete one but ignore the structural relationships in local regions. They are less competent in learning point distributions and recovering the details of the object. This paper proposes a shape-aware point cloud completion network (SCNet) that employs multi-scale features and a coarse-to-fine strategy to generate detailed, complete point clouds. Firstly, we introduce a K-feature nearest neighbor algorithm to explore local geometric structure and design a novel shape-aware graph convolution that utilizes multiple learnable filters to perceive local shape changes in different directions. Secondly, we adopt non-local feature expansion to generate a coarse point cloud as the rough shape and merge it with the input data to preserve the original structure. Finally, we employ a residual network to fine-tune the point coordinates to smooth the merged point cloud, which is then optimized to a fine point cloud using a refinement module with shape-aware graph convolution and local attention mechanisms. Extensive experiments demonstrate that our SCNet outperforms other methods on the same point cloud completion benchmark and is more stable and robust.

Layout generation is an emerging computer vision task that incorporates the challenges of object localization and aesthetic evaluation, widely used in advertisements, posters, and slides design. An ideal layout should consider both the intra-domain relationship within layout elements and the inter-domain relationship between layout elements and the image. However, most previous methods simply focus on image-content-agnostic layout generation without leveraging the complex visual information from the image. To address this limitation, we propose a novel paradigm called image-conditioned layout generation, which aims to add text overlays to an image in a semantically coherent manner. Specifically, we introduce the Image-Conditioned Variational Transformer (ICVT) that autoregressively generates diverse layouts in an image. Firstly, the self-attention mechanism is adopted to model the contextual relationship within layout elements, while the cross-attention mechanism is used to fuse the visual information of conditional images. Subsequently, we take them as building blocks of the conditional variational autoencoder (CVAE), which demonstrates attractive diversity. Secondly, to alleviate the gap between the layout elements domain and the visual domain, we design a Geometry Alignment module, in which the geometric information of the image is aligned with the layout representation. Thirdly, we present a self-refinement mechanism to automatically refine the failure case of generated layout, effectively improving the quality of generation. Experimental results show that our model can adaptively generate layouts in the non-intrusive area of the image, resulting in a harmonious layout design.

Camouflaged object detection (COD) is a challenging task due to its irregular shape and color similarity or even blending into the surrounding environment. It is difficult to achieve satisfactory results by directly using salient object detection methods due to the low contrast with the surrounding environment and obscure object boundary in camouflaged object detection. To determine the location of the camouflaged objects and achieve accurate segmentation, the interaction between features is essential. Similarly, an effective feature aggregation method is also very important. In this paper, we propose a contextual fusion and feature refinement network (CFNet). Specifically, we propose a multiple-receptive-fields-based feature extraction module (MFM) that obtains features from multiple scales of receptive fields. Then, the features are input to an attention-based information interaction module (AIM), which establishes the information flow between adjacent layers through an attention mechanism. Finally, the features are fused and optimized layer by layer using a feature fusion module (FFM). We validate the proposed CFNet as an effective COD model on four benchmark datasets, and the generalization ability of our proposed model is verified in the salient object detection task.

The proliferation of ultra-high-definition (UHD) imaging device is increasingly being used for underwater image acquisition. However, due to light scattering and underwater impurities, UHD underwater images often suffer from color deviations and edge blurriness. Many studies have attempted to enhance underwater images by integrating frequency domain and spatial domain information. Nonetheless, these approaches often interactively fuse dual-domain features only in the final fusion module, neglecting the complementary and guiding roles of frequency domain and spatial domain features. Additionally, the extraction of dual-domain features is independent of each other, which leads to the sharp advantages and disadvantages of the dual-domain features extracted by these methods. Consequently, these methods impose high demands on the feature fusion capabilities of the fusion module. But in order to handle UHD underwater images, the fusion modules in these methods often stack only a limited number of convolution and activation function operations. This limitation results in insufficient fusion capability, leading to defects in the restoration of edges and colors in the images. To address these issues, we develop a dual-domain interaction network for enhancing UHD underwater images. The network takes into account both frequency domain and spatial domain features to complement and guide each other’s feature extraction patterns, and fully integrates the dual-domain features in the model to better recover image details and colors. Specifically, the network consists of a U-shaped structure, where each layer is composed of dual-domain interaction transformer blocks containing interactive multi-head attention and interactive simple gate feed-forward networks. The interactive multi-head attention captures local interaction features of frequency domain and spatial domain information using convolution operation, followed by multi-head attention operation to extract global information of the mixed features. The interactive simple gate feed-forward network further enhances the model’s dual-domain interaction capability and cross-dimensional feature extraction ability, resulting in clearer edges and more realistic colors in the images. Experimental results demonstrate that the performance of our proposal in enhancing underwater images is significantly better than existing methods.

Due to the rapid spread of rumors on social media, which has a detrimental effect on our lives, it is becoming increasingly important to detect rumors. It has been proved that the study of dynamic graphs is helpful to capture the temporal change of information transmission and understand the evolution trend and pattern change of events. However, the dynamic learning methods currently studied do not fully consider the interaction characteristics of the evolutionary process. Therefore, it is difficult to fully capture the structural and semantic differences between them. In order to fully exploit the potential correlations of such temporal information, we propose a novel model named dynamic evolution characteristics learning (DECL) method for rumor detection. First, we partition the temporal snapshot sequences based on the propagation structure of rumors. Secondly, a multi-task graph contrastive learning method is adopted to enable the graph encoder to capture the essential features of rumors, and to fully explore the temporal structural differences and semantic similarities between true rumor and false rumor events. Experimental results on three real-world social media datasets confirm the effectiveness of our model for rumor detection tasks.

Differential evolution (DE) is a cutting-edge meta-heuristic algorithm known for its simplicity and low computational overhead. But the traditional DE cannot effectively balance between exploration and exploitation. To solve this problem, in this paper, a dynamic dual-population DE variant (ADPDE) is proposed. Firstly, the dynamic population division mechanism based on individual potential value is presented to divide the population into two subgroups, effectively improving the population diversity. Secondly, a nonlinear reduction mechanism is designed to dynamically adjust the size of potential subgroup to allocate computing resources reasonably. Thirdly, two unique mutation strategies are adopted for two subgroups respectively to better utilise the effective information of potential individuals and ensure fast convergence speed. Finally, adaptive parameter setting methods of two subgroups further achieve the balance between exploration and exploitation. The effectiveness of improved strategies is verified on 21 classical benchmark functions. Then, to verify the overall performance of ADPDE, it is compared with three standard DE algorithms, eight excellent DE variants and seven advanced evolutionary algorithms on CEC2013, CEC2017 and CEC2020 test suites, respectively, and the results show that ADPDE has higher accuracy and faster convergence speed. Furthermore, ADPDE is compared with eight well-known optimizers and CEC2020 winner algorithms on nine real-world engineering optimization problems, and the results indicate ADPDE has the development potential for constrained optimization problems as well.

The accuracy of stock price forecasting is of great significance in investment decision-making and risk management. However, the complexity and fluctuation of stock prices challenge the traditional forecasting methods to achieve the best accuracy. To improve the accuracy of stock price prediction, a sophisticated combination prediction method based on ICEEMDAN-FA-BiLSTM–GM has been proposed in this article. In this paper, a comprehensive and effective indicator system is constructed, covering 60 indicators such as traditional factors, market sentiment, macroeconomic indicators and company financial data, which affect stock prices. In the data preprocessing stage, in order to eliminate the influence of noise, the stock closing price series is first decomposed by using the ICEEMDAN method, which effectively divides them into high-frequency and low-frequency components according to their respective frequencies. Subsequently, LLE technique is used to narrow down the remaining indicators to obtain 9 narrowed features. Finally, each high-frequency subsequence is combined with all the dimensionality reduction features respectively to construct new indicator sets for input to the model. In the prediction stage, the hyperparameters of the prediction model for each subseries have been determined using the FA algorithm. The prediction has been carried out separately for the high-frequency and low-frequency components, employing the BiLSTM and GM prediction methods. Ultimately, the prediction results of each subseries have been superimposed to obtain the final stock price prediction value. In this paper, an empirical study was conducted using stock price data such as Shanghai composite index. The experimental results show that the established stock price prediction model based on ICEEMDAN-FA-BiLSTM–GM has obvious advantages in terms of prediction accuracy and stability compared with traditional methods and other combined prediction methods. This model can provide more accurate stock price prediction and promote the rationalization of investment decision and the accuracy of risk control.

Digital image forensics, particularly in the realm of detecting Copy-Move Forgery (CMF), is exposed to significant challenges, especially in the face of intricate adversarial attacks. In response to these challenges, this paper presents a robust approach for detecting complex CMFs in digital images using the KeyPoint-Siamese Capsule Network (KP-SCN) and evaluates its resilience against adversarial attacks. The KP-SCN architecture incorporates keypoint detection, a Siamese network for feature extraction, and a capsule network for forgery detection. The method showcases enhanced robustness against adversarial attacks, specifically addressing image perturbation, patch removal, patch replacement, and spatial transformation attacks. By using hierarchical feature representations and dynamic routing in capsule networks, the model effectively handles complex CMF, including rotation, scaling, and non-linear transformations. The proposed KP-SCN approach employs a large dataset for training the KP-SCN, enabling it to identify copy-move forgeries by comparing extracted keypoints and their spatial relationships. KP-SCN demonstrates superior performance compared to the state-of-the-art on the CoMoFoD dataset, achieving precision, recall, and F1-score values of 95.62%, 93.78%, and 94.69%, respectively, and shows strong results on other datasets. For CASIA v2.0, the precision, recall, and F1-score are 90.45%, 88.97%, and 89.70%; for MICC-F2000, they are 91.32%, 90.27%, and 90.79%; for MICC-F600, they are 92.21%, 91.10%, and 91.65%; for MICC-F8multi, they are 89.75%, 87.92%, and 88.83%; and for IMD, they are 93.14%, 92.58%, and 92.86%. The KP-SCN framework maintains high detection rates under various manipulations, including JPEG compression, rotation, scaling, noise, blurring, brightness changes, contrast adjustment, and zoom motion blur compared to the other methods. For instance, it achieves an 80.657% detection rate for CoMoFoD under JPEG compression and 97.883% for IMD under a 10-degree rotation. These findings validate the robustness and adaptability of KP-SCN, making it a reliable solution for real-world forensic applications.