Jailbreak attacks pose a significant threat to the reliable deployment of large language models (LLMs) in critical applications. Although existing LLMs are supervised fine-tuning and aligned through reinforcement learning from human feedback, automated jailbreak attack algorithms can still identify potential jailbreak prompts that lead to harmful outputs. In this paper, we propose JPS, a jailbreak attack detector based on selective perturbation and contrastive learning. JPS leverages the robustness of jailbreak attacks, which is achieved through complex multi-step optimization, by using perturbation methods to enhance the training data for jailbreak prompts. In order to mitigate noise from perturbation, we introduce a selective strategy based on token importance. Additionally, we employ supervised contrastive learning to effectively differentiate between jailbreak and benign samples. Extensive experiments on the popular jailbreak attacks and benign datasets show that JPS outperforms all the baseline approaches according to F1-score. Furthermore, detailed ablation experiments were conducted to analyze each module of the model, demonstrating the effectiveness of our approach.

With the booming rise of the Internet and social media, semantically rich multimodal data has gradually become the mainstream carrier of news dissemination. Among them, multi-modal fake news with illustrations and text has attracted widespread attention due to its greater deceptiveness. However, existing research methods are mainly limited to the analysis of images and text within the news itself, failing to fully consider the consistency and discrepancy characteristics between different modalities, which hinders the full exploitation of the advantages of multi-modal fusion. To address this issue, this study proposes a multi-modal fake news detection method with large language models(LLMs) and Knowledge-Aligned Attention Networks(MFDnet). This method first leverages the powerful semantic understanding capabilities of large language models to generate detailed text descriptions for images, serving as a knowledge supplement for the image model. Subsequently, by constructing a Knowledge-Aligned Attention Networks, it achieves efficient semantic fusion between the knowledge-supplemented image model and text modal information, thereby effectively extracting the consistency and complementary features between different modalities. Experimental results demonstrate that this model exhibits excellent performance on multiple public fake news detection datasets.

Generating high-quality synonymous perturbations is a core challenge for textual adversarial tasks. However, candidates generated from the masked language model often contain many words that are antonyms or irrelevant to the original words, which limit the perturbation space and affect the attack’s effectiveness. We present ProAttacker1which uses Prompt to make the mask language models better adversarial Attackers. ProAttacker inverts the prompt paradigm by leveraging the prompt with the class label to guide the language model to generate more semantically-consistent perturbations. We present a systematic evaluation to analyze the attack performance on 6 NLP datasets, covering text classification and inference. Our experiments demonstrate that ProAttacker outperforms state-of-the-art attack strategies in both success rate and perturb rate.

Automated security screening has a significant role In protecting public spaces from security threats by employing X-ray images to detect prohibited items. However, there are challenges of noise production due to squeezing, occlusion, and penetration of luggage objects. Additionally, the hues of objects are monotonous and lack luster. To solve these problems, we propose an Autonomous Baggage Threat Detection Network (ABTD-Net) for accurate prohibited item detection. To tackle the difficulty of capturing distinctive visual features, we constructed a Feature Adjustment Head (FAH) to refine pyramid features. Specifically, we designed an Attention Module (AM) at several places after initially using a Dense Unidirectional Propagation (DUP) to filter noise. Furthermore, we created a Feature Fusion Head (FFH) that dynamically fuses hierarchical visual information under object occlusion, including early-fusion and late-fusion. Extensive experiments on security inspection X-ray datasets OPIXray and HiXray demonstrate the superiority of our proposed method.

Logs are widely used during the development and maintenance of software systems. Logs assist developers and operation & maintenance personnel to understand the state and behavior of systems at runtime. Also, logs can diagnose system failures and conduct abnormal analyses to provide further protection to the security of systems. However, large software systems generate large amounts of semi-structured logging routinely. The first step to support further analysis is how to parse semi-structured records with free-form text log messages into structured templates. Therefore, log parsing is rather challenging. Because logs are generated by static templates (i.e., log statements) in the source code, templates are often not accessible when parsing logs. It is worth noting that most proposed approaches still rely on log-specific heuristics or manual rule extraction. Those existed methods are often specialized for parsing certain log types and often neglect the semantic meaning of log messages, thus limiting performance scores and generalization, hence, in this paper, we propose a new parsing technique - Adaptive Contextual Aware Attention Network for Log Parsing via Word Classification, named AdaptParse. Adapt-Parse transforms the template generation problem into a word classification task, then learns the features of template words and variable words. We evaluate our AdaptParse on 5 realworld log datasets and compare the performance with 7 parsing techniques. Our experimental results show that the proposed approach can effectively understand the semantic meaning of log messages and achieve accurate log parsing results. Overall, AdaptParse achieves state-of-the-art performance on five realworld log datasets, outperforming all the baseline models.

Anomaly detection of key performance indicators (KPIs), e.g., CPU load, network usage, is crucial for system behavior monitoring. In recent years, several anomaly detection approaches have been proposed. However, detecting anomalies of KPIs remains challenging because of the stochastic nature and complex temporal dependence of multivariate time series. Additionally, the presence of noise and the unavailability of labeled data in large-scale datasets limit the effectiveness of anomaly detection. In this paper, we propose IAD-Net, an interpretable anomaly detection method with Dual Gated Residual Fusion Networks. The main idea is to model the inter-metric and temporal dependencies simultaneously by using gated residual blocks and two-stream fusion. Additionally, we utilize Gated Recurrent Unit (GRU) to extract long-term global trend patterns of an input sequence. Finally, both the forecasting-based model and the reconstruction-based model are combined in order to focus on single-timestamp predictions and latent representations of time series. Extensive experiments on real-world data show that IAD-Net outperforms other state-of-the-art approaches according to F1-score. Further analysis confirms the effectiveness of our method in anomaly interpretation.

Community question answering (CQA) becomes increasingly prevalent in recent years, providing platforms for users with various backgrounds to obtain information and share knowledge. However, the redundancy and lengthiness issues of crowd-sourced answers limit the performance of answer selection, thus leading to difficulties in reading or even misunderstandings for community users. To solve these problems, we propose the dual graph question-answer attention networks (DGQAN) for answer selection task. Aims to fully understand the internal structure of the question and the corresponding answer, firstly, we construct a dual-CQA concept graph with graph convolution networks using the original question and answer text. Specifically, our CQA concept graph exploits the correlation information between question-answer pairs to construct two sub-graphs (QSubject-Answer and QBody-Answer), respectively. Further, a novel dual attention mechanism is incorporated to model both the internal and external semantic relations among questions and answers. More importantly, we conduct experiment to investigate the impact of each layer in the BERT model. The experimental results show that DGQAN model achieves state-of-the-art performance on three datasets (SemEval-2015, 2016, and 2017), outperforming all the baseline models.

Malicious traffic detection is an important task in network security, which protects the target network from privacy leakage and service paralysis. The complexity of the network and the hierarchical structure of network traffic, i.e, byte-packet-flow, indicate the diversity of traffic information. Most of the existing work only uses one feature or statistical feature, and cannot learn network traffic from multiple perspectives, i.e, shortsighted, which results in the lack of important information in network traffic. Meanwhile, after obtaining multiple features, the effective fusion of multiple features is also an urgent problem to be solved. In this paper, we propose a Multiple Features Fusion with Attention Networks (MFFAN). According to the hierarchical structure of network traffic, we extract byte, packet, and statistical features from original traffic files to learn traffic from multiple perspectives, overcoming shortsighted. To effectively fuse multiple features, we use the self-attention to learn the intra-feature relationship with each feature and use the co-attention to learn the inter-feature relationship between features. We conduct experiments on the ISCIDS2012 dataset and CICIDS2017 dataset, and the results show that our model achieves an effective fusion of multiple features and high accuracy.

In recent years, sarcasm detection on social media has become one of the major challenges in natural language processing, due to the figurative characteristics of sarcasm. With the rise of social media platforms, especially Twitter, which allows users to post textual content while attaching a complement (such as image and video). A prevailing trend is to achieve a satire effect via combinations of text and image on Twitter. Therefore, to overcome the above-mentioned difficulty of sarcasm detection on social media, the key is to establish a multi-modal framework that can synchronically capture the information contained in text and vision. In this research, we propose Coupled-Attention Networks (CANs), which can effectively integrate information of text and image into a unified framework, thus realizing the fusion of different forms of resources. We conduct our experiments on a real-world dataset. Experimental results prove that our method achieves excellent results.

The objective and universal nature of user behavior data make it the primary data for insider threat detection. Existing solutions treat user behavior as atomic symbols and do not consider behavior semantic information. Meanwhile, fine-grained temporal information is ignored despite its relevance to describe user behavior. Such approaches inevitably lead to unsatisfactory performance and generalization. In this paper, we propose ITDBERT which embeds temporal information into behavior and catches the fused semantic representation via pre-trained language models. ITDBERT also leverages attention-based Bi-LSTM to provide behavior-level detection results. To verify the effectiveness of our proposed method, we conduct comparison experiments on Cert datasets. Our proposed model achieves an F1-score of 0.9243 in day-level insider threat detection, which outperforms baselines.

With the rapid development of computer technology, the traditional Internet data security and information privacy issues are gradually expanding to all aspects of society as a whole. As the first line of defense for information security, identity authentication technology becomes crucial. Among the many authentication technologies, continuous authentication technology has gained increasing attention. In this paper, we design fusion keystroke time-textual attention networks for continuous authentication based on the keystroke data (keystroke time series, keystroke text) when users enter free-text. Specifically, the corresponding keystroke time series and the corresponding keystroke text are first obtained based on the original keystroke data, and then the keystroke time series and the keystroke text are input into the BiLSTM model and the pre-training model, respectively; the BiLSTM can better capture the temporal features, and the pre-training model can better capture the textual features when authenticating the user. Finally, the two information are fed into the cross attention model to better integrate the two information. Experiments show that the FKTAN model achieves promising results on two datasets, Clarkson II keystroke dataset and Buffalo dataset, outperforming all baseline models.

Most of the existed fake news detection works on social media driven-fake news mainly focused on text. However, more and more social media platforms like Twitter, facebook, etc, allow users to create multi-modal contents, including text, image and video. Hence, it is obvious that only investigating text contents is insufficient to achieve solid detection. In this paper, we study the fake news on social media platforms composed of multimodal contents (text and images), and propose Dual Attention Fusion Networks for fake news detection on social media. We explore three modalities, (text modality, image modality and image attributes modality), and further propose a Dual Attention Fusion Networks (DAFN) model for this task. First, our proposed model extracts text modality and image modality, respectively. We then pass combinations of image attributes modality and text modality through BERT to extract text features. Finally, we reconstruct features of three modalities and fuse them into a feature vector for prediction. Our method is verified on realworld datasets consisting of collected social media platforms. Experiments show that the our method achieves promising results on real world datasets. outperforming all baseline models.