ABSTRACT This paper presents an efficient GAN‐based adversarial attack method to spoof deepfake detectors. While such detectors, built on deep neural networks, show high accuracy in identifying forgeries, they are vulnerable to adversarial perturbations. Our proposed model employs a two‐branch architecture: one branch generates general, image‐independent perturbations, while the other enhances the adversarial efficacy of the reconstructed images. Through joint training, we generate visually realistic outputs that severely degrade detector performance. Experiments on FaceForensics++ demonstrate the effectiveness of the proposed method. It achieves competitive performance by reducing detection accuracy below 16% while maintaining high visual fidelity. The resulting highly imperceptible adversarial samples highlight a significant vulnerability in existing detectors.

Abstract Network Intrusion Detection Systems (NIDS) are essential for safeguarding networks against malicious activities. However, existing machine learning-based NIDS often require complex feature engineering, which demands significant domain expertise and experimentation, leading to suboptimal model performance in complex network environments. In contrast, deep learning approaches, while powerful, struggle with imbalanced data, resulting in a bias towards normal traffic and reduced effectiveness in detecting rare attacks. To address these issues, we propose a method that combines contrastive learning and Bayesian Gaussian Mixture Model (BGMM). Specifically, we propose a novel contrastive learning loss that enables the model to automatically learn the similarity within normal traffic and the distinction between normal and malicious traffic, thereby generating robust and distinguishable feature representations. This approach not only eliminates the need for manual feature engineering but also helps alleviate the issue of weak feature representations for rare attacks. BGMM further enhances detection performance by adapting to both normal and malicious patterns through the use of multiple components. The effectiveness of the proposed method is validated through extensive experiments on two widely used modern network intrusion datasets. On the UNSW-NB15 dataset, the proposed method achieves 91.27% accuracy and 92.30% F1-score, which is 1.85% and 2.35% better than the state-of-the-art (SOTA) method. On the Distrinet-CIC-IDS2017 dataset, the proposed method achieves 99.66% accuracy and 99.12% F1-score, which is 0.05% and 0.12% better than the SOTA method.

Abstract Graph Neural Networks (GNNs) have shown excellent performance as a powerful tool on link prediction task. Recent studies have shown that link prediction based on GNNs is vulnerable to backdoor attacks. However, existing backdoor attack methods on link prediction task require modification of the link state, which results in poor stealthiness of the backdoor. To address this issue, a clean-label backdoor attack method on link prediction task (CL-Link) is proposed in this paper. Specifically, CL-Link utilizes subgraphs as backdoor triggers and achieves trigger injection by attaching subgraphs to target links. In order to enhance the stealthiness of the attack, CL-Link attaches the trigger without modifying the original connection state of the target links. Instead, it utilizes the original connection state as the label, thus minimizing disturbances to the dataset. To ensure the effectiveness of the attack, the gradient information of the model and the similarity between the trigger nodes and the nodes in the graph are used to optimize the features of the trigger nodes. Extensive experiments were performed on multiple benchmark datasets (i.e., Cora, Citeseer, and Pubmed), and the proposed method achieved the highest attack success rate of 97.69% with a poisoning rate of only 5%, which validates the effectiveness of our proposed approach.

Abstract Data collection is an effective way to build a better Graph Neural Network (GNN) model, but it also makes it easy for attackers to implant backdoors into the model through data poisoning. In this work, we propose a backdoor detection method of graph for data collection scenarios (CGBD). Different from most existing backdoor detection methods of Neural Network (NN) models, especially the Deep Neural Network (DNN) models, the difference in predictions of backdoor samples in clean and backdoor models is exploited for backdoor detection in CGBD. Specifically, in the backdoor model, the backdoor samples with modified labels are predicted as the target class. However, in the clean model, they are predicted as the ground-truth labels since the clean model remains unaffected by the backdoor. Due to the detection methodology of CGBD is not based on the potential forms of triggers, it can detect backdoor samples with any type of trigger. Additionally, since data is associated with its providers, CGBD can detect not only backdoor data but also malicious data providers. Extensive experiments on multiple benchmark datasets demonstrate that data with varying poisoning rates exhibit significant anomalies compared to clean data. This validates the effectiveness of our proposed method.

Graph Neural Network(GNN) have gained extensive adoption in diverse fields, including IoT anomaly detection, social network analysis, and drug molecule prediction, due to their exceptional ability to handle graph-structured data. However, as GNN models become more widely used, the issue of dataset leakage has become increasingly prominent, posing significant risks to the rights and interests of data owners. In this paper, we propose a data ownership authentication method based on GNN backdoor watermarking, termed Graph Data Ownership Authentication(GDOA). Specifically, the data owner injects covert backdoor watermark triggers into some of the samples in the dataset. When an unauthorized user uses the data to train their own model, the model will be injected with the backdoor. The data owner can then validate the target model using samples with the same triggers to determine whether the target model was trained using the unauthorized dataset. GDOA utilizes the clean-label backdoor method to achieve ownership authentication of the dataset due to the greater stealthiness of the clean-label backdoor. Specifically, the watermark samples are selected within the target class, and the samples’ own labels are used as target labels, effectively avoiding the issue of label confusion. In addition, to improve the authentication success rate, GDOA optimizes the injection position of feature triggers in the feature vector. Our extensive experiments across multiple models and benchmark datasets demonstrate that GDOA achieves an average authentication success rate of over 90%, validating its effectiveness for graph data ownership verification.

warm systems have attracted growing interest due to Internet of Things (IoT) advancements and their applications in multi-robot control, environmental mapping, and intelligent transportation. Average consensus protocols are crucial in these scenarios, but current implementations require frequent interagent data exchange, posing significant risks to network security. warm systems have attracted growing interest due to Internet of Things (IoT) advancements and their applications in multi-robot control, environmental mapping, and intelligent transportation. Average consensus protocols are crucial in these scenarios, but current implementations require frequent interagent data exchange, posing significant risks to network security. In this paper, we present a privacy-preserving consensus algorithm designed to ensure the privacy of the initial state while achieving consensus on the precise average of the initial values, even in the presence of denial-of-service (DoS) attacks. First, to protect agent state privacy, we introduce perturbation to the transmitted information among the agents using independently generated random noises drawn from Laplace distributions, in conjunction with the utilization of specifically designed pseudo-random values. Then, by synthesizing the perturbed average consensus algorithm with an event-triggered control strategy, a distributed privacy-preserving consensus protocol is established for the swarm systems under DoS attacks. The presented results indicate that under the given network communication conditions, the event-triggered mechanism can effectively withstand the impact of DoS attacks on the network and ensure the achievement of accurate average consensus. Finally, experimental results on a UAV-swarm board based Wi-Fi network are provided to validate the effectiveness of obtained theoretical results.

The proliferation of facial manipulation has been propelled by generative adversarial networks (GAN), severely threatening to the personal privacy and reputation. Accordingly, one such countermeasure is adversarial watermark, which is embedded into the protected image prior to GAN synthesization attack, resulting into the distorted fake content obtained by malicious attackers. However, in practice, JPEG compression usually causes a remarkable degradation on the performance of adversarial watermark. To address this challengeable issue, this letter presents a novel robust adversarial watermark, which can effectively defend against GAN synthesization attack, even though suffering from JPEG compression. Extensive experiments verify the superiority of our proposed method in the benchmark dataset; more importantly, the robustness of the proposed adversarial watermark is comprehensively evaluated on the both simulated transmission channel and the realism social network platform.

This article investigates the secure and privacy-preserving consensus problem of multiagent systems (MASs) with directed interaction topologies under multiple cyberattacks, which contain deception attacks and DoS attacks. First, a unified attack model is introduced to characterize such a multiple attack phenomenon. Besides, considering the existence of eavesdroppers who can intercept the data transmitted on the links, a fully distributed agent value reconstruction method based on the idea of state decomposition is designed to prevent the leakage of the agent's initial information. Then, a novel privacy-preserving adaptive resilient consensus algorithm (PPARCA) with certain graph robustness condition for MASs under the multiple cyberattacks is proposed. The algorithm adaptively takes different countermeasures in the face of different cyberattacks. PPARCA uses the reconstructed agents' states and combines with the modified secure acceptance and broadcast algorithm (SABA). Theoretical analysis shows that the proposed algorithm can effectively protect the privacy of the initial state of the agents, and reach resilient consensus in the face of cyberattacks. Finally, numerical simulations and Raspberry Pi MASs practical application experiments demonstrate the effectiveness of the proposed results.

With the development of facial recognition technology, face anti-spoofing as the most important security module of face recognition system becomes more and more important. As a matter of fact, face anti-spoofing is still a challenging task, especially facing various attacks simultaneously. Moreover, most of current detectors mainly focus on binary classification while usually fail to complete the task of fine-grained multiple classification, referring to as replay, print, partial mask, and full mask attacks. To fill the gap, in this context, it is proposed to design the fine-grained detection network for classifying various face spoofing attack modes. First, we propose to establish a Transformer style network structure for feature extraction, where the convolution mapping operation is adopted instead of traditional linear mapping. Specifically, we adopt the self-attention module for extracting long distance feature, and convolution mapping is used to maintain the model's ability to extract local features. Finally, the simple yet effective linear classifier is introduced for fine-grained classification. Moreover, with the help of the VGG based style-transfer network, the well-designed scheme of data augmentation module is proposed for solving the problem of insufficient training samples. In the large-scale experiments, compared with the baseline detectors, our proposed fine-grained classifier with low computation cost performs its superiority for multiple classification.

With the continuous development of computer hardware equipment and deep learning technology, it is easier for people to swap faces in videos by currently-emerging multimedia tampering tools, such as the most popular deepfake. It would bring a series of new threats of security. Although many forensic researches have focused on this new type of manipulation and achieved high detection accuracy, most of which are based on supervised learning mechanism with requiring a large number of labeled samples for training. In this paper, we first develop a novel unsupervised detection manner for identifying deepfake videos. The main fundamental behind our proposed method is that the face region in the real video is taken by the camera while its counterpart in the deepfake video is usually generated by the computer; the provenance of two videos is totally different. Specifically, our method includes two clustering stages based on Photo-Response Non-Uniformity (PRNU) and noiseprint feature. Firstly, the PRNU fingerprint of each video frame is extracted, which is used to cluster the full-size identical source video (regardless of its real or fake). Secondly, we extract the noiseprint from the face region of the video, which is used to identify (re-cluster for the task of binary classification) the deepfake sample in each cluster. Numerical experiments verify our proposed unsupervised method performs very well on our own dataset and the benchmark FF++ dataset. More importantly, its performance rivals that of the supervised-based state-of-the-art detectors.

Most existing face authentication systems have limitations when facing the challenge raised by presentation attacks, which probably leads to some dangerous activities when using facial unlocking for smart device, facial access to control system, and face scan payment. Accordingly, as a security guarantee to prevent the face authentication from being attacked, the study of face presentation attack detection is developed in this community. In this work, a face presentation attack detector is designed based on residual color texture representation (RCTR). Existing methods lack of effective data preprocessing, and we propose to adopt DW-filter for obtaining residual image, which can effectively improve the detection efficiency. Subsequently, powerful CM texture descriptor is introduced, which performs better than widely used descriptors such as LBP or LPQ. Additionally, representative texture features are extracted from not only RGB space but also more discriminative color spaces such as HSV, YCbCr, and CIE 1976 L∗a∗b (LAB). Meanwhile, the RCTR is fed into the well-designed classifier. Specifically, we compare and analyze the performance of advanced classifiers, among which an ensemble classifier based on a probabilistic voting decision is our optimal choice. Extensive experimental results empirically verify the proposed face presentation attack detector’s superior performance both in the cases of intradataset and interdataset (mismatched training-testing samples) evaluation.

In current steganalysis, relying on a large scale of samples, widely-adopted supervised schemes require the training stage while few studies focus on the design of a training-free unsupervised adaptive detector with high efficiency. To fill the gap, we investigate an adaptive statistical model-based detector designed for detecting JPEG steganography. First, in virtue of hypothesis testing theory, together with the distribution of quantized DCT coefficients, we establish the general framework of the statistical model-based detector. Second, based on the framework, we mainly analyze the performance of the detector relying on the selection of the statistical model, parameters estimation, and less significant payload prediction. Third, to improve the reliability of detection, based on the strategy of assigning weights for DCT channels, the novel adaptive statistical model-based detectors are proposed to aim at detecting JPEG steganography, involving the channel-selected or non-channel-selected algorithm. Extensive experiments highlight the effectiveness of the proposed methodology. Moreover, when detecting JPEG images adopted by two steganographic schemes with the small payload, the experimental results show the Area Under Curve (AUC) of our proposed optimal adaptive detector can achieve as high as 0.9567 and 0.9895 respectively, which are both better than that of non-adaptive detector.

Security of consensus control is of key significance in multi-agent systems. In this paper we investigate the resilient consensus problem for multi-agent systems under the specific attack scenarios where the attacker can eavesdrop on initial information of agents among the system, and modifies the values in the communication links to interfere in the consensus process. To protect the privacy of node' value, we employ the cryptography of homomorphic encryption to encrypt initial integer state of each agent, without revealing to other agents the real value in the network. When the communication network meets the necessary connectivity, we develop a variation of so-called ratio consensus algorithm that deal with malicious attacks. The numerical example of directed network is conducted to illustrate the effectiveness of our proposed algorithm.

Group Recommendation means recommending satisfying activities to a group of users in social networks. Most existing aggregation methods mainly focus on aggregating the user's rating for the item according to the rating feature. They ignore the rationality and interpretability of user weights in aggregate function. Moreover, the user weights determined by these methods cannot be adaptively changed with different groups and items. In this paper, we propose an adaptive (AG) aggregation method based on item genre. The AG method does two things: first, uses item genre information to extract a reasonable user weight; second, performs adaptive weight modeling. In addition, we consider that there will be interactions among members in the group decision-making process, so the weights of members may be adjusted appropriately. In order to consider this reasonable factor in our recommendation system, we improved the AG method from the perspective of group member weight deviation. Extensive experimental results on MovieLens dataset clearly demonstrate the effectiveness of our proposed AG.

Point of interest (POI) recommendation has become an important means to help people discover attractive locations. Previous studies show that modeling the context information of users' check-ins is necessary for POI recommendation. However, most previous hybrid models ignore the fact that a user's interested region is likely to drive him to visit other POIs in it, which means that his region-level interest is indispensable in the recommendation process. In this paper, we take this challenge and propose a hybrid model called POI-Region POI Recommendation (PRPOIR). Specifically, PRPOIR contains two modules: Interest Module and Context Module. In Interest Module, inspired by the success of the Logistic Matrix Factorization (LMF) to model implicit feedback, we apply it to model users' POI-level interest and region-level interest respectively. To capture the difference in the users' emphasis on region-level interests, we innovatively introduce a strategy to establish the dynamic weight of users' region-level interest. Then in Context Module, by further integrating geographical influence and social correlations into PRPOIR, it can alleviate the sparsity of check-ins and bring further performance improvement. To verify the effectiveness of PRPOIR, we conduct experiments on two real-world datasets, experimental results show that PRPOIR outperforms the state-of-the-art models.

Passenger-searching strategies, as the crowd intelligence of massive taxi drivers, are hidden in their historical GPS traces. Mining traces to understand the efficient passenger searching strategies can benefit drivers themselves. Traditionally, the research on passenger search strategies from taxi GPS traces is mainly focused on statistical techniques. Although this can improve the ability of drivers to find potential passengers through hotspots recommendation, most of these research still directly use raw GPS data and failed to take drivers' experience into account. Moreover, because driver's experience is behind of raw data, can't obtain directly, so the traditional model is unable to make good use of it during hot spots mining. In this paper, we proposed an inference model based on HITS (Hypertext Induced Topic Search), which perfectly describes the relationship between hot spots and drivers' experience and thus effectively handle the above problem. We first extract hotspots by an innovative PDBSCAN algorithm based on fuzzy grid partition and match them with corresponding landmarks by a landmark matching algorithm which based on kernel density estimation. Then the HITS-based inference model is used to mine popular hotspots and the most experienced drivers. Finally, we plan an optimal path for drivers. Experimental results demonstrate the efficiency and the ability of this method to provide drivers with better hotspots and hunting sequences recommendation.

The widespread of location-based social networks has generated massive check-in sequences in chronological order. Forecasting check-in sequences is significant while challenging due to the check-ins' sparsity problem. Existing methods have followed closely to incorporate spatial and temporal context to alleviate the data sparsity problem, but neglect the semantic sequential correlation between check-ins. Howbeit, incorporating the semantic sequential correlation between check-ins for next POI recommendation encounters the challenges of semantic sequential correlation measurement and sequential behavior modeling. To measure the semantic sequential correlation, we apply a semantic sequential correlation calculation model based on a semantic correlational graph that incorporates the time intervals' influence to calculate the semantic sequential correlation. Then, we apply a novel Long Short-Term Memory (LSTM) framework equipped with two additional semantic gates that takes the additional semantic sequential correlation as the extra input to capture users' sequential behaviors and model their long short-term interest with the restrictions in the semantic level. Finally, we cluster users into different groups as an improvement of our model to achieve a more accurate recommendation. Our proposed model is evaluated on a real-world and large-scale dataset and the experimental results demonstrate that our method outperforms the state-of-the-art methods for next POI recommendation.

The problem of authenticating a re-sampled image has been investigated over many years. Currently, however, little research proposes a statistical model-based test, resulting in that statistical performance of the resampling detector could not be completely analyzed. To fill the gap, we utilize a parametric model to expose the traces of resampling forgery, which is described with the distribution of residual noise. Afterward, we propose a statistical model describing the residual noise from a resampled image. Then, the detection problem is cast into the framework of hypothesis testing theory. By considering the image content with designing a texture weight map, two types of statistical detectors are established. In an ideal context in which all distribution parameters are perfectly known, the likelihood ratio test (LRT) is presented and its performance is theoretically established. An upper bound of the detection power can be successfully obtained from the statistical performance of an LRT. For practical use, when the distribution parameters are not known, a generalized LRT with three different maps based on estimation of parameters is established. Numerical results on simulated data and real natural images highlight the relevance of our proposed approach.

User Identification problem is concerned with identifying the same person with multiple virtual identities across social network sites(SNSs). Most of the existing approaches pays close attention to the similarity of profile attributes, generate-contents and linkages of friends or simply combination of these features. Only one method analyzes the feasibility of user tags in User Identification problems, but does not analyze the particularity and the inconsistency of tags belong to users among different social networks. In this paper, an improved user identification method across social networks via tagging behaviors is proposed that a new symmetric variant of BM25 (BM25 is a bag-of-words retrieval function that ranks a set of documents based on the query terms appearing in each document, regardless of the inter-relationship between the query terms within a document) using the semantic relationships between inconsistent tags among different social networks. By using extracted features from the inconsistent tagging behaviors, profile attributes and SVM supervised learning techniques, a classifier is developed for performing user identity matching between two social network sites. Evaluation on Douban and Weibo real world data-set showed that the accuracy of the proposed method is 30% higher than that of the common tag-based approach.

Most bike-sharing service providers offer a free ride for a short time period. In this paper, we study how to find the optimal route that is free of rental cost and minimizes the trip distance from one location to another within a bike-sharing system, in which the utilization of bike stations dynamically changes over time. We use a time-dependent dynamic graph to model the network of bike stations. In the graph, each vertex represents a bike station and is associated with a vertex-usage function. The difficulty of this problem is mainly attributed to the fluctuation of the usage function because a fully utilized station cannot accept returned bikes. Efficiency is another challenge, as we must explore all possible paths between the source and the destination. To address these challenges, we propose techniques to find a solution optimized for efficiency. First, to reduce the search space, we construct a station network graph on top of a road network. Next, we employ a pathstree to identify all the paths with lengths less than the userpreferred maximum detour distance and we select an optimal path toward the destination. We present the design details of our algorithms and we analyze the algorithms' correctness and complexity. To demonstrate the feasibility of our methods, we also report the results from the extensive experiments we conducted.

The network traffic classification is one of the most fundamental work in the network measurement and management, and this problem is more and more impact as the network scale grows. Many methods are proposed by researchers, but methods based on flow statistics seem more popular than the others. In this paper, we proposed a novel method based on refined flow statistical features. The new statistics, skewness and kurtosis, and new flow statistical features, payload length, were introduced into raw feature set firstly. Then, with the consideration of efficiency in the classification stage, the feature selection was used on the raw feature set to get an optimal feature set and the feature selection are mainly based on the K-means clustering algorithm. The comparison experiment results show that the proposed optimal feature set reaches the same precision level with half time consuming and internal cluster distance when compared with the raw set.