A recommender system is susceptible to manipulation through the injection of carefully crafted profiles. Some recent profile identification methods only perform well in specific attack scenarios. A general attack detection method is usually complicated or requires label samples. Such methods are prone to overtraining easily, and the process of annotation incurs high expenses. This study proposes an unsupervised divide-and-conquer method aiming to identify attack profiles, utilizing a specifically designed model for each kind of shilling attack. Initially, our method categorizes the profile set into two attack types, namely Standard and Obfuscated Behavior Attacks. Subsequently, profiles are separated into clusters within the extracted feature space based on the identified attack type. The selection of attack profiles is then determined through target item analysis within the suspected cluster. Notably, our method offers the advantage of requiring no prior knowledge or annotation. Furthermore, the precision is heightened as the identification method is designed to a specific attack type, employing a less complicated model. The outstanding performance of our model, validated through experimental results on MovieLens-100K and Netflix under various attack settings, demonstrates superior accuracy and reduced running time compared to current detection methods in identifying Standard and Obfuscated Behavior Attacks.

Clothes changing is one of the challenges in person re-identification (ReID), since clothes provide remarkable and reliable information for decision, especially when the resolution of an image is low. Variation of clothes significantly downgrades standard ReID models, since the clothes information dominates the decisions. The performance of the existing methods considering clothes changing is still not satisfying, since they fail to extract sufficient identity information that excludes clothes information. This study aims to disentangle identity, clothes, and unrelated features with a Generative Adversarial Network (GAN). A GAN model with three encoders, one generator, and three discriminators, and its training procedure are proposed to learn these kinds of features separately and exclusively. Experimental results indicate that our model generally achieves the best performance among state-of-the-art methods in both ReID tasks with and without clothes changing, which confirms that the identity, clothes, and unrelated features are extracted by our model more precisely and effectively.

Surface electromyogram (sEMG)-based hand gesture recognition, which interprets commands given by humans through sEMG signals, performs well in many studies. However, its recognition accuracy drops dramatically due to electrode shift since the distributions of motion classes are changed. Although calibrating the system with newly collected samples after electrode shift maintains the accuracy, collecting labeled samples is inconvenient and time-consuming since the procedure is rigid. However, the calibration may not work properly without label, especially when the change is significant. This study proposes a user friendly and convenient calibration method for hand gesture recognition by an unsupervised domain adaptation method, which only obtains the unlabeled samples of preselected benchmark classes from users in calibration. The change of benchmark classes is captured by unlabeled samples by a clustering method. The other classes are estimated based on the benchmark classes by regression models. As a result, the information of all classes is used to calibrate the system. Linear discriminant analysis is used to demonstrate our model. A dataset with ten subjects is collected to verify the performance empirically. Experimental results confirm that our method utilizes the unlabeled benchmark class samples in calibration and achieves 75.55% average accuracy. Our method is more robust to electrode shift and improves around 8.5% accuracy consistently on all subjects compared with the methods without calibration or label information in calibration. Although the accuracy of our method is slightly less than the ones using label calibration samples, our calibration data collection is more convenient and less complicated.

In a modern e-commerce recommender system, it is important to understand the relationships among products. Recognizing product relationships-such as complements or substitutes-accurately is an essential task for generating better recommendation results, as well as improving explainability in recommendation. Products and their associated relationships naturally form a product graph, yet existing efforts do not fully exploit the product graph's topological structure. They usually only consider the information from directly connected products. In fact, the connectivity of products a few hops away also contains rich semantics and could be utilized for improved relationship prediction. In this work, we formulate the problem as a multilabel link prediction task and propose a novel graph neural network-based framework, item relationship graph neural network (IRGNN), for discovering multiple complex relationships simultaneously. We incorporate multihop relationships of products by recursively updating node embeddings using the messages from their neighbors. An edge relational network is designed to effectively capture relational information between products. Extensive experiments are conducted on real-world product data, validating the effectiveness of IRGNN, especially on large and sparse product graphs.

Feature selection plays an important role in machine learning in order to reduce model complexity and extract more meaningful information. The recent studies indicate that not only the generalization ability but also the security should be considered in selecting features in an adversarial environment in which a classifier may be misled by an adversary intentionally. However, since only the nearest legitimate sample to a malicious sample is considered, the existing adversarial filter feature selection method is sensitive to outlier samples and is not suitable for Boolean features. A distribution-based adversarial filter feature selection method is proposed against evasion attack in this study. Our method uses distribution-based measurements, such as Symmetric Uncertainty and Earth Mover's Distance, are used to quantify the generalization ability and security of a feature subset. The experiments suggest our proposed method outperforms existing methods in terms of robustness and stability in datasets with Boolean and real-valued features.

Image-level weakly supervised semantic segmentation (WSSS) reduces the cost of semantic segmentation significantly as only category labels are required. In the WSSS pipeline, the initially obtained seed areas are imprecise and should be refined. AffinityNet is a widely used seed area refinement method which learns pixel-level affinity between coordinates to adjust the object regions. However, the capabilities of AffinityNet are limited due to the complex simultaneous exploration of the affinities of all object classes, as the classes have different characteristics. This paper proposes a class-specific AffinityNet (CSANet) in which the affinities of each class are learnt separately by the class-specific affinity layers. Moreover, as the learning task is simple because the affinity of only one class is focused, the information contained by the uncertain neutral regions is extracted to provide additional information in the training of the class-specific affinity layers. The empirical results demonstrate the effectiveness of our method, which captures more precise boundaries and significantly improves segmentation results on the PASCAL VOC 2012 segmentation benchmark.

Security problems of deep reinforcement learning draw much attention recently. Previous works on adversary attack mainly focus on preventing the targeted agent from choosing the most desirable action at each step, which may not reduce the cumulative reward effectively. In this paper, we first investigate how changing features affect the cumulative reward achieved by an agent. The static reward impact map is introduced to quantify the influence on the reward of each feature experimentally. By focusing on tasks with the static reward impact map, an adversarial attack method against deep reinforcement learning aiming to minimize the cumulative reward is proposed. Features with the large reward impact are perturbed in crafting an adversarial sample. Deep Q-network is selected to demonstrate the performance of our attack method in the experiments. The results indicate that our proposed method achieves better performance than the existing one-time attack method and the random attack in terms of the cumulative reward and the successful attack rate under both white-box and black-box settings.

Convolutional Neural Network (CNN) achieved satisfying performance in click-through rate (CTR) prediction in recent studies. Since features used in CTR prediction have no meaningful sequence in nature, the features can be arranged in any order. As CNN learns the local information of a sample, the feature sequence may influence its performance significantly. However, this problem has not been fully investigated. This paper firstly investigates whether and how the feature sequence affects the performance of the CNN-based CTR prediction method. As the data distribution of CTR prediction changes with time, the best current sequence may not be suitable for future data. Two multi-sequence models are proposed to learn the information provided by different sequences. The first model learns all sequences using a single feature learning module, while each sequence is learnt individually by a feature learning module in the second one. Moreover, a method of generating a set of embedding sequences which aims to consider the combined influence of all feature pairs on feature learning is also introduced. The experiments are conducted to demonstrate the effectiveness and stability of our proposed models in the offline and online environment on both the benchmark Avazu dataset and a real commercial dataset.

Collaborative filtering (CF) is vulnerable under shilling attack, which misleads recommendation of CF by injecting well-crafted profiles to a targeted system. Although a number of supervised learning based shilling attack detection methods are proposed, their features mainly measure rating values and items of a profile individually, but ignore the relation between items. This study aims to enhance the robustness of CF against shilling attack by considering the rated item correlation. Real users rate items based on their preferences, but rated items are randomly selected for malicious users profiles in most shilling attack. Therefore, the rated item correlation of real and malicious profiles is different. Three features are proposed to capture the information from different intervals of the distribution of rated item correlation in terms of Cosine Association (CA). A benchmark dataset, MovieLens 100K, is used to evaluate the proposed features. The discrimination ability of the proposed features is also illustrated. The experimental results suggest that the proposed features have significant contribution on shilling attack detection.

Face recognition technique has been widely applied to personal identification systems due to its satisfying performance. However, its security may be a crucial issue, since many studies have shown that face recognition systems may be vulnerable in an adversarial environment, in which an adversary can camouflage as a legitimate user in order to mislead the system. Although face liveness detection methods have been proposed to distinguish real and fake faces, they are either time-consuming, costly, or sensitive to noise and illumination. This paper proposes a face liveness detection method with flash against 2D spoofing attack. Flash not only can enhance the differentiation between legitimate and illegitimate users, but it also reduces the influence of environmental factors. Two images are taken from a subject, one with flash and another without flash. Four texture and 2D structure descriptors with low computational complexity are used to capture information of the two images in our model. Advantages of our method include low installation cost of flash and no user cooperation required. A data set of 50 subjects collected under different scenarios is used in the experiments to evaluate the proposed method. The experimental results indicate that the proposed model performs better than existing liveness detection methods in different environmental scenarios. This paper confirms that the use of flash successfully improves face liveness detection in terms of accuracy, robustness, and running time.

Pattern recognition and machine learning techniques have been increasingly adopted in adversarial settings such as spam, intrusion, and malware detection, although their security against well-crafted attacks that aim to evade detection by manipulating data at test time has not yet been thoroughly assessed. While previous work has been mainly focused on devising adversary-aware classification algorithms to counter evasion attempts, only few authors have considered the impact of using reduced feature sets on classifier security against the same attacks. An interesting, preliminary result is that classifier security to evasion may be even worsened by the application of feature selection. In this paper, we provide a more detailed investigation of this aspect, shedding some light on the security properties of feature selection against evasion attacks. Inspired by previous work on adversary-aware classifiers, we propose a novel adversary-aware feature selection model that can improve classifier security against evasion attacks, by incorporating specific assumptions on the adversary's data manipulation strategy. We focus on an efficient, wrapper-based implementation of our approach, and experimentally validate its soundness on different application examples, including spam and malware detection.

The training of a multilayer perceptron neural network (MLPNN) concerns the selection of its architecture and the connection weights via the minimization of both the training error and a penalty term. Different penalty terms have been proposed to control the smoothness of the MLPNN for better generalization capability. However, controlling its smoothness using, for instance, the norm of weights or the Vapnik-Chervonenkis dimension cannot distinguish individual MLPNNs with the same number of free parameters or the same norm. In this paper, to enhance generalization capabilities, we propose a stochastic sensitivity measure (ST-SM) to realize a new penalty term for MLPNN training. The ST-SM determines the expectation of the squared output differences between the training samples and the unseen samples located within their Q -neighborhoods for a given MLPNN. It provides a direct measurement of the MLPNNs output fluctuations, i.e., smoothness. We adopt a two-phase Pareto-based multiobjective training algorithm for minimizing both the training error and the ST-SM as biobjective functions. Experiments on 20 UCI data sets show that the MLPNNs trained by the proposed algorithm yield better accuracies on testing data than several recent and classical MLPNN training methods.

This paper addresses a problem in the hashing technique for large scale image retrieval: learn a compact hash code to reduce the storage cost with performance comparable to that of the long hash code. A longer hash code yields a better precision rate of retrieved images. However, it also requires a larger storage, which limits the number of stored images. Current hashing methods employ the same code length for both queries and stored images. We propose a new hashing scheme using two hash codes with different lengths for queries and stored images, i.e., the asymmetric cyclical hashing. A compact hash code is used to reduce the storage requirement, while a long hash code is used for the query image. The image retrieval is performed by computing the Hamming distance of the long hash code of the query and the cyclically concatenated compact hash code of the stored image to yield a high precision and recall rate. Experiments on benchmarking databases consisting up to one million images show the effectiveness of the proposed method.

Steganography detection is a technique to tell whether there are secret messages hidden in images. The performance of a steganalysis system is mainly determined by the method of feature extraction and the architecture selection of the classifier. Selecting a proper classifier with proper parameters will improve the detection accuracy and generalization capability of the system. We propose a Radial Basis Function Neural Network (RBFNN) optimized by the Localized Generalization Error Model (L-GEM) for steganograhpy detection. In the proposed method, the discrete cosine transform (DCT) features and the Markov features are used as inputs of neural networks for detection. To enhance the generalization capability of the RBFNN and the performance of detecting steganography in future images, the architecture of the RBFNN is selected by minimizing the L-GEM. The experimental results show that the proposed method provides a better performance on testing images in comparison with the existing method in attackting Steghide, OutGuess and F5.

A new dynamic classifier fusion method named L-GEM fusion method (LFM) for multiple classifier systems (MCSs) is proposed. The localized generalization error upper bound for the neighborhood of a testing sample is calculated and used to estimate the local competence of base classifiers in MCSs. Different from the recent dynamic classifier selection methods, the proposed method consider not only the training error but also the sensitivity of the base classifier. Experimental results show that the MCSs using LFM has more accurate than other popular dynamic fusion methods.

Multiple Classifier System (MCS) is a very popular research topic in recent years. It has been proved theoretically and empirically to be better than single classifiers in many scenarios. Creating diverse sets of classifier is one of the key issues in building MCSs. Feature grouping is one of the methods to create diverse classifiers and it has been shown to improve the accuracy of an MCS. In this paper, we propose a new feature grouping method based on Genetic Algorithm (GA) with the localized Generalization Error Model as the evaluation criterion. The combined individual classifiers using the weighted sum are examined in this paper. Moreover, several feature grouping methods are compared with the proposed method in this work. The experimental results on benchmark dataset show that the MCS trained by the proposed method is promising.

Multiple Classifier Systems (MCSs) have been shown theoretically and empirically to outperform a single classifier in many applications. However, many ensemble training algorithms sometimes create a very large MCS which is combined by many individual classifiers. A large MCS not only consumes computational resources but also decreases the effectiveness. One of the solutions is the pruning method. It reduces the number of individual classifiers inside an MCS that maintains the performance well or is just slightly worse than the original one. In this paper, a new pruning method, called NNEPSM, for Neural Network ensemble based on a sensitivity measure is proposed. The classifiers which have less impact to the final output of MCS will be removed. The advantages of this method include efficient performance, low-complexity and independence on training method. NNEPSM has been applied in Web applications and other benchmark dataset. The experimental results showed that our approach performs well using different datasets.

Large margin classifiers have been widely applied in solving supervised learning problems. One representative model in large margin learning is the support vector machine (SVM). SVM is an unstructured classifier since the data structure information is underutilized and the decision hyperplane calculation relies exclusively on the support vectors. To incorporate the data covariance information into the large margin learning, structured large margin machine (SLMM) is recently proposed and show better performance than classical SVM in some applications. Instead of utilizing the data structures straightly like SLMM, SVM ensemble (SVMe) improves the generalization ability of SVM in another way by combining the outputs of a series of SVMs. Inspired by SVMe, we are going to explore the ensemble counterpart for SLMM, i.e., SLMMe, and validate the effectiveness of multiple SLMM system. Experimental results on benchmark datasets demonstrate that SLMMe improves SLMM by reducing its variance, and SLMMe outperforms SVMe in most cases in terms of both classification accuracy and variance.

In classification problem, the learning process can be more efficient if the informative samples can be selected actively based on the knowledge of the classifier. This problem is called active learning. Most of the existing active learning methods did not directly relate to the generalization error of classifiers. Also, some of them need high computational time or are based on strict assumptions. This paper describes a new active learning strategy using the concept of localized generalization error of the candidate samples. The sample which yields the largest generalization error will be chosen for query. This method can be applied to different kinds of classifiers and its complexity is low. Experimental results demonstrate that the prediction accuracy of the classifier can be improved by using this selecting method and fewer training samples are possible for the same prediction accuracy.