## 2023 Vol. 32, No. 1

column
Display Method:
2023, 32(1): 1-12. doi: 10.23919/cje.2022.00.031
Abstract(569) HTML (280) PDF(136)
Abstract:
Collaborative inference (co-inference) accelerates deep neural network inference via extracting representations at the device and making predictions at the edge server, which however might disclose the sensitive information about private attributes of users (e.g., race). Although many privacy-preserving mechanisms on co-inference have been proposed to eliminate privacy concerns, privacy leakage of sensitive attributes might still happen during inference. In this paper, we explore privacy leakage against the privacy-preserving co-inference by decoding the uploaded representations into a vulnerable form. We propose a novel attack framework named AttrLeaks, which consists of the shadow model of feature extractor (FE), the susceptibility reconstruction decoder, and the private attribute classifier. Based on our observation that values in inner layers of FE (internal representation) are more sensitive to attack, the shadow model is proposed to simulate the FE of the victim in the black-box scenario and generates the internal representations. Then, the susceptibility reconstruction decoder is designed to transform the uploaded representations of the victim into the vulnerable form, which enables the malicious classifier to easily predict the private attributes. Extensive experimental results demonstrate that AttrLeaks outperforms the state of the art in terms of attack success rate.
2023, 32(1): 13-25. doi: 10.23919/cje.2021.00.312
Abstract(256) HTML (127) PDF(38)
Abstract:
2023, 32(1): 26-38. doi: 10.23919/cje.2022.00.038
Abstract(459) HTML (221) PDF(51)
Abstract:
Malware detection has been a hot spot in cyberspace security and academic research. We investigate the correlation between the opcode features of malicious samples and perform feature extraction, selection and fusion by filtering redundant features, thus alleviating the dimensional disaster problem and achieving efficient identification of malware families for proper classification. Malware authors use obfuscation technology to generate a large number of malware variants, which imposes a heavy analysis burden on security researchers and consumes a lot of resources in both time and space. To this end, we propose the MalFSM framework. Through the feature selection method, we reduce the 735 opcode features contained in the Kaggle dataset to 16, and then fuse on metadata features (count of file lines and file size) for a total of 18 features, and find that the machine learning classification is efficient and high accuracy. We analyzed the correlation between the opcode features of malicious samples and interpreted the selected features. Our comprehensive experiments show that the highest classification accuracy of MalFSM can reach up to 98.6% and the classification time is only 7.76 s on the Kaggle malware dataset of Microsoft.
2023, 32(1): 39-50. doi: 10.23919/cje.2021.00.276
Abstract(257) HTML (130) PDF(29)
Abstract:
A software ecosystem (SECO) can be described as a special complex network. Previous complex networks in an SECO have limitations in accurately reflecting the similarity between each pair of nodes. The community structure is critical towards understanding the network topology and function. Many scholars tend to adopt evolutionary optimization methods for community detection. The information adopted in previous optimization models for community detection is incomprehensive and cannot be directly applied to the problem of community detection in an SECO. Based on this, a complex network in SECOs is first built. In the network, the cooperation intensity between developers is accurately calculated, and the attribute contained by each developer is considered. A multi-objective optimization model is formulated. A community detection algorithm based on NSGA-II is employed to solve the above model. Experimental results demonstrate that the proposed method of calculating the developer cooperation intensity and our model are advantageous.
2023, 32(1): 51-63. doi: 10.23919/cje.2022.00.044
Abstract(422) HTML (209) PDF(83)
Abstract:
Many advanced object detection algorithms are mainly based on natural scenes object and rarely dedicated to fine-grained objects. This seriously limits the application of these advanced detection algorithms in remote sensing object detection. How to apply horizontal detection in remote sensing images has important research significance. The mainstream remote sensing object detection algorithms achieve this task by angle regression, but the periodicity of angle leads to very large losses in this regression method, which increases the difficulty of model learning. Circular smooth label (CSL) solved this problem well by transforming the regression of angle into a classification form. YOLOv5 combines many excellent modules and methods in recent years, which greatly improves the detection accuracy of small objects. We use YOLOv5 as a baseline and combine the CSL method to learn the angle of arbitrarily oriented targets, and distinguish the fine-grained between instance classes by adding an attention mechanism module to accomplish the fine-grained target detection task for remote sensing images. Our improved model achieves an average category accuracy of 39.2% on the FAIR1M dataset. Although our method does not achieve satisfactory results, this approach is very efficient and simple, reducing the hardware requirements of the model.
2023, 32(1): 64-75. doi: 10.23919/cje.2022.00.077
Abstract(283) HTML (138) PDF(39)
Abstract:
Shui manuscripts are part of the national intangible cultural heritage of China. Owing to the particularity of text reading, the level of informatization and intelligence in the protection of Shui manuscript culture is not adequate. To address this issue, this study created Shuishu_C, the largest image dataset of Shui manuscript characters that has been reported. Furthermore, after extensive experimental validation, we proposed ShuiNet-A, a lightweight artificial neural network model based on the attention mechanism, which combines channel and spatial dimensions to extract key features and finally recognize Shui manuscript characters. The effectiveness and stability of ShuiNet-A were verified through multiple sets of experiments. Our results showed that, on the Shui manuscript dataset with 113 categories, the accuracy of ShuiNet-A was 99.8%, which is 1.5% higher than those of similar studies. The proposed model could contribute to the classification accuracy and protection of ancient Shui manuscript characters.
2023, 32(1): 76-84. doi: 10.23919/cje.2022.00.009
Abstract(240) HTML (121) PDF(27)
Abstract:
Deep learning based language models have improved generation-based linguistic steganography, posing a huge challenge for linguistic steganalysis. The existing neural-network-based linguistic steganalysis methods are incompetent to deal with complicated text because they only extract single-granularity features such as global or local text features. To fuse multi-granularity text features, we present a novel linguistic steganalysis method based on attentional bidirectional long-short-term-memory (BiLSTM) and short-cut dense convolutional neural network (CNN). The BiLSTM equipped with the scaled dot-product attention mechanism is used to capture the long dependency representations of the input sentence. The CNN with the short-cut and dense connection is exploited to extract sufficient local semantic features from the word embedding matrix. We connect two structures in parallel, concatenate the long dependency representations and the local semantic features, and classify the stego and cover texts. The results of comparative experiments demonstrate that the proposed method is superior to the state-of-the-art linguistic steganalysis.
2023, 32(1): 85-96. doi: 10.23919/cje.2021.00.093
Abstract(318) HTML (152) PDF(28)
Abstract:
Few-shot learning (FSL) is a new machine learning method that applies the prior knowledge from some different domains tasks. The existing FSL models of metric-based learning have some drawbacks, such as the extracted features cannot reflect the true data distribution and the generalization ability is weak. In order to solve the problem in the present, we developed a model named cross modal adaptive few-shot learning based on task dependence (COOPERATE for short). A feature extraction and task representation method based on task condition network and auxiliary co-training is proposed. Semantic representation is added to each task by combining both visual and textual features. The measurement scale is adjusted to change the property of parameter update of the algorithm. The experimental results show that the COOPERATE has the better performance comparing with all approaches of the monomode and modal alignment FSL.
2023, 32(1): 97-110. doi: 10.23919/cje.2022.00.034
Abstract(167) HTML (83) PDF(20)
Abstract:
To solve the rating prediction problems of low accuracy and data sparsity on different datasets, we propose an interactive perception method based collaborative rating prediction algorithm named DCAE-MF, by fusing dual convolutional autoencoder (DCAE) and probability matrix factorization (PMF). Deep latent representations of users and items are captured simultaneously by DCAE and are deeply integrated with PMF to collaboratively make rating predictions based on the known rating history of users. A global multi-angle collaborative optimization learning method is developed to effectively optimize all the parameters of DCAE-MF. Extensive experiments are performed on seven real-world datasets to demonstrate the superiority of DCAE-MF on key rating accuracy metrics of the root mean squared error (RMSE) and mean absolute error (MAE).
2023, 32(1): 111-129. doi: 10.23919/cje.2022.00.100
Abstract(221) HTML (111) PDF(20)
Abstract:
A large number of metaheuristics have been proposed and shown high performance in solving complex optimization problems. While most variation operators in existing metaheuristics are empirically designed, new operators are automatically designed in this work, which are expected to be search space independent and thus exhibit robust performance on different problems. This work first investigates the influence of translation invariance, scale invariance, and rotation invariance on the search behavior and performance of some representative operators. This work then deduces the generic form of translation, scale, and rotation invariant operators, and proposes a principled approach for the automated design of operators, which searches for high-performance operators based on the deduced generic form. The experimental results demonstrate that the operators generated by the proposed approach outperform state-of-the-art ones on a variety of problems with complex landscapes and up to 1000 decision variables.
2023, 32(1): 130-139. doi: 10.23919/cje.2021.00.122
Abstract(465) HTML (208) PDF(43)
Abstract:
Samples collected from most industrial processes have two challenges: one is contaminated by the non-Gaussian noise, and the other is gradually obsolesced. This feature can obviously reduce the accuracy and generalization of models. To handle these challenges, a novel method, named the robust online extreme learning machine (RO-ELM), is proposed in this paper, in which the least mean $\boldsymbol{p}$-power criterion is employed as the cost function which is to boost the robustness of the ELM, and the forgetting mechanism is introduced to discard the obsolescence samples. To investigate the performance of the RO-ELM, experiments on artificial and real-world datasets with the non-Gaussian noise are performed, and the datasets are from regression or classification problems. Results show that the RO-ELM is more robust than the ELM, the online sequential ELM (OS-ELM) and the OS-ELM with forgetting mechanism (FOS-ELM). The accuracy and generalization of the RO-ELM models are better than those of other models for online learning.
2023, 32(1): 140-150. doi: 10.23919/cje.2022.00.096
Abstract(267) HTML (131) PDF(39)
Abstract:
We present graph Hilbert neural network (GHNN), a novel framework of graph neural networks based on graph signal processing theory, which is different from the previous method based on convolution theorem. Graph Hilbert transform (GHT) can explain the emergence of complex eigenvalues and complex eigenvectors, which provides a theoretical basis for convolution operation on digraphs. GHT can be expressed by the form of polynomial filter because of its linear shift invariant (LSI) property and the definition in spectral domain, which is applied to construct the layers of graph neural network. The graph Laplacian matrix is adopted as the graph shift operator in the function of LSI filter to realize the property of localization. To make better use of both low-frequency and high-frequency information, we design a two-channel filter bank to perform low-pass filtering and high-pass filtering simultaneously. Experiments on three benchmark datasets show that the proposed GHNN outperforms previous spectral graph convolutional neural network on the task of graph-based semi-supervised classification.
2023, 32(1): 140-150. doi: 10.23919/cje.2022.00.040
Abstract(119) HTML (65) PDF(5)
Abstract:
Tongue color is an important tongue diagnostic index for traditional Chinese medicine (TCM). Due to the individual experience of TCM experts as well as ambiguous boundaries among the tongue color categories, there often exist noisy labels in annotated samples. Deep neural networks trained with the noisy labeled samples often have poor generalization capability because they easily overfit on noisy labels. A novel framework named confident-learning-assisted knowledge distillation (CLA-KD) is proposed for tongue color classification with noisy labels. In this framework, the teacher network plays two important roles. On the one hand, it performs confident learning to identify, cleanse and correct noisy labels. On the other hand, it learns the knowledge from the clean labels, which will then be transferred to the student network to guide its training. Moreover, we elaborately design a teacher network in an ensemble manner, named E-CA2-ResNet18, to solve the unreliability and instability problem resulted from the insufficient data samples. E-CA2-ResNet18 adopts ResNet18 as the backbone, and integrates channel attention (CA) mechanism and activate or not activation function together, which facilitates to yield a better performance. The experimental results on three self-established TCM tongue datasets demonstrate that, our proposed CLA-KD can obtain a superior classification accuracy and good robustness with a lower network model complexity, reaching 94.49%, 92.21%, 93.43% on the three tongue image datasets, respectively.
2023, 32(1): 151-158. doi: 10.23919/cje.2021.00.309
Abstract(438) HTML (194) PDF(60)
Abstract:
The robustness of adversarial examples to image scaling transformation is usually ignored when most existing adversarial attacks are proposed. In contrast, image scaling is often the first step of the model to transfer various sizes of input images into fixed ones. We evaluate the impact of image scaling on the robustness of adversarial examples applied to image classification tasks. We set up an image scaling system to provide a basis for robustness evaluation and conduct experiments in different situations to explore the relationship between image scaling and the robustness of adversarial examples. Experiment results show that various scaling algorithms have a similar impact on the robustness of adversarial examples, but the scaling ratio significantly impacts it.
2023, 32(1): 159-165. doi: 10.23919/cje.2019.00.028
Abstract(243) HTML (122) PDF(29)
Abstract:
For more effective image sampling, compressive sensing (CS) imaging methods based on image saliency have been proposed in recent years. Those methods assign higher measurement rates to salient regions, but lower measurement rate to non-salient regions to improve the performance of CS imaging. However, those methods are block-based, which are difficult to apply to actual CS sampling, as each photodiode should strictly correspond to a block of the scene. In our work, we propose a non-uniform CS imaging method based on image saliency, which assigns higher measurement density to salient regions and lower density to non-salient regions, where measurement density is the number of pixels measured in a unit size. As the dimension of the signal is reduced, the quality of reconstructed image will be improved theoretically, which is confirmed by our experiments. Since the scene is sampled as a whole, our method can be easily applied to actual CS sampling. To verify the feasibility of our approach, we design and implement a hardware sampling system, which can apply our non-uniform sampling method to obtain measurements and reconstruct the images. To our best knowledge, this is the first CS hardware sampling system based on image saliency.
2023, 32(1): 166-176. doi: 10.23919/cje.2021.00.138
Abstract(215) HTML (110) PDF(17)
Abstract:
Synthetic aperture radar (SAR) imaging is an efficient strategy which exploits the properties of microwaves to capture images. A major concern in SAR imaging is the reconstruction of image from back scattered signals in the presence of noise. The reflected signal consist of more noise than the target signal and it is a challenging problem to reduce the noise in the collected signal for better reconstruction of an image. Current studies mostly focus on filtering techniques for noise removal. This can result in an undesirable point spread function causing extreme smearing effect in the desired image. In order to handle this problem, a computational technique, particle swarm optimization (PSO) is used for de-noising purpose and later the target performance is further improved by an amalgamation of Wiener filter. Moreover, to improve the de-noising performance we have exploited the singular value decomposition based morphological filtering. To justify the proposed improvements we have simulated the proposed techniques and results are compared with the conventional existing models. The proposed method revealed considerable decrease in mean square error compared to Wiener filter and PSO techniques. Quantitative analysis of image restoration quality are also presented in comparison with Wiener filter and PSO based on the improvement in signal to noise ratio and peak signal to noise ratio.
2023, 32(1): 177-188. doi: 10.23919/cje.2021.00.084
Abstract(219) HTML (108) PDF(29)
Abstract:
The key to multi-sensor image fusion is the fusion of infrared and visible images. Fusion of infrared and visible images with generative adversarial network (GAN) has great advantages in automatic feature extraction and subjective vision improvement. Due to different principle between infrared and visible imaging, the blur phenomenon of edge and texture is caused in the fusion result of GAN. For this purpose, this paper conducts a novel generative adversarial network with blur suppression. Specifically, the generator uses the residual-in-residual dense block with switchable normalization layer as the elemental network block to retain the infrared intensity and the fused image textural details and avoid fusion artifacts. Furthermore, we design an anti-blur loss function based on Weber local descriptor. Finally, numerous experiments are performed qualitatively and quantitatively on public datasets. Results justify that the proposed method can be used to produce a fusion image with sharp edge and clear texture.
2023, 32(1): 189-198. doi: 10.23919/cje.2021.00.211
Abstract(246) HTML (125) PDF(16)
Abstract:
Real-time six degrees-of-freedom (6D) object pose estimation is essential for many real-world applications, such as robotic grasping and augmented reality. To achieve an accurate object pose estimation from RGB images in real-time, we propose an effective and lightweight model, namely high-resolution 6D pose estimation network (HRPose). We adopt the efficient and small HRNetV2-W18 as a feature extractor to reduce computational burdens while generating accurate 6D poses. With only 33% of the model size and lower computational costs, our HRPose achieves comparable performance compared with state-of-the-art models. Moreover, by transferring knowledge from a large model to our proposed HRPose through output and feature-similarity distillations, the performance of our HRPose is improved in effectiveness and efficiency. Numerical experiments on the widely-used benchmark LINEMOD demonstrate the superiority of our proposed HRPose against state-of-the-art methods.