Extreme events frequently occur in real-world time series and often carry significant practical implications. In domains such as climate and healthcare, these events, such as floods, heatwaves, or acute medical episodes, can lead to serious consequences. Accurate forecasting of such events is therefore of substantial importance. Most existing time series forecasting models are optimized for overall performance within the prediction window, but often struggle to accurately predict extreme events, such as high temperatures or heart rate spikes. The main challenges are data imbalance and the neglect of valuable information contained in intermediate events that precede extreme events. In this paper, we propose xTime, a novel framework for extreme event forecasting in time series. xTime leverages knowledge distillation to transfer information from models trained on lower-rarity events, thereby improving prediction performance on rarer ones. In addition, we introduce a MoE mechanism that dynamically selects and fuses outputs from expert models across different rarity levels, which further improves the forecasting performance for extreme events. Experiments on multiple datasets show that xTime achieves consistent improvements, with forecasting accuracy on extreme events improving from 3% to 78%.

Graphs are pervasive in the real-world, such as social network analysis, bioinformatics, and knowledge graphs. Graph neural networks (GNNs) have great ability in node classification but still face challenges in scenarios with few labeled nodes, despite the frequent occurrence of such tasks in real-world applications with limited labeled data. To address this challenge, various approaches have been proposed, including graph meta-learning, transfer learning, and methods based on Large Language Models (LLMs). However, traditional meta-learning and transfer learning methods often require prior knowledge from base classes or fail to exploit the potential advantages of unlabeled nodes. Meanwhile, LLM-based methods may overlook the zero-shot capabilities of LLMs and rely heavily on the quality of generated contexts. In this paper, we propose a novel approach that integrates LLMs and GNNs, leveraging the zero-shot inference capabilities of LLMs and employing a Graph-LLM-based active learning paradigm to enhance GNNs’ performance. Extensive experiments demonstrate the effectiveness of our model in improving node classification accuracy with considerably limited labeled data, surpassing state-of-the-art baselines by significant margins.

Social media has drastically reshaped the world that allows billions of people to engage in such interactive environments to conveniently create and share content with the public. Among them, text data (e.g., tweets, blogs) maintains the basic yet important social activities and generates a rich source of user-oriented information. While those explicit sensitive user data like credentials have been significantly protected by all means, personal private attribute (e.g., age, gender, location) disclosure due to inference attacks is somehow challenging to avoid, especially when powerful natural language processing (NLP) techniques have been effectively deployed to automate attribute inferences from implicit text data. This puts users’ attribute privacy at risk. To address this challenge, in this article, we leverage the inherent vulnerability of machine learning to adversarial attacks, and design a novel text-space Adv ersarial attack for S ocial G ood, called Adv4SG . In other words, we cast the problem of protecting personal attribute privacy as an adversarial attack formulation problem over the social media text data to defend against NLP-based attribute inference attacks. More specifically, Adv4SG proceeds with a sequence of word perturbations under given constraints such that the probed attribute cannot be identified correctly. Different from the prior works, we advance Adv4SG by considering social media property, and introducing cost-effective mechanisms to expedite attribute obfuscation over text data under the black-box setting. Extensive experiments on real-world social media datasets have demonstrated that our method can effectively degrade the inference accuracy with less computational cost over different attribute settings, which substantially helps mitigate the impacts of inference attacks and thus achieve high performance in user attribute privacy protection.

As malicious bots reside in a network to disrupt network stability, graph neural networks (GNNs) have emerged as one of the most popular bot detection methods. However, in most cases these graphs are significantly class-imbalanced. To address this issue, graph oversampling has recently been proposed to synthesize nodes and edges, which still suffers from graph heterophily, leading to suboptimal performance. In this paper, we propose HOVER, which implements Homophilic Oversampling Via Edge Removal for bot detection on graphs. Instead of oversampling nodes and edges within initial graph structure, HOVER designs a simple edge removal method with heuristic criteria to mitigate heterophily and learn distinguishable node embeddings, which are then used to oversample minority bots to generate a balanced class distribution without edge synthesis. Experiments on TON IoT networks demonstrate the state-of-the-art performance of HOVER on bot detection with high graph heterophily and extreme class imbalance.

Graphs have emerged as one of the most important and powerful data structures to perform content analysis in many fields. In this line of work, node classification is a classic task, which is generally performed using graph neural networks (GNNs). Unfortunately, regular GNNs cannot be well generalized into the real-world application scenario when the labeled nodes are few. To address this challenge, we propose a novel few-shot node classification model that leverages pseudo-labeling with graph active learning. We first provide a theoretical analysis to argue that extra unlabeled data benefit few-shot classification. Inspired by this, our model proceeds by performing multi-level data augmentation with consistency and contrastive regularizations for better semi-supervised pseudo-labeling, and further devising graph active learning to facilitate pseudo-label selection and improve model effectiveness. Extensive experiments on four public citation networks have demonstrated that our model can effectively improve node classification accuracy with considerably few labeled data, which significantly outperforms all state-of-the-art baselines by large margins.

To mitigate evolving malware attacks, machine learning models have been successfully deployed to detect malware. However, these models are often challenged by data scarcity, design efforts and constrained resources. Inspired by the adversarial vulnerability of machine learning, in this paper, we design a novel model Adv4Mal to adversarially reprogram an ImageNet classification neural network for malware detection in both white-box and black-box settings. As such, a small or moderate amount of data are sufficient to train a promising malware detection model, the varying software features can be uniformly processed without extra efforts, and the majority of computation can be wisely shared and reused to save the resources. This, to the best of our knowledge, has not yet been explored. Specifically, Adv4Mal proceeds by embedding software features into a host image to construct new data, and learning a universal perturbation to be added to all inputs in an imperceptible manner, such that the outputs of the pretrained model can be accordingly mapped to the final detection decisions for all software. We evaluate Adv4Mal on three software datasets. The experimental results demonstrate that Adv4Mal can successfully exploit ImageNet model's learning capability and limited data to achieve high performance in malware detection, and also yield significant advantages of model flexibility to different features, and cost efficiency in computing resources.

The popularization of social media generates a large amount of user-oriented data, where text data especially attracts researchers and speculators to infer user attributes (e.g., age, gender) for fulfilling their intents. Generally, this line of work casts attribute inference as a text classification problem, and starts to leverage graph neural networks for higher-level text representations. However, these text graphs are constructed on words, suffering from high memory consumption and ineffectiveness on few labeled texts. To address this challenge, we design a text-graph-based few-shot learning model for social media attribute inferences. Our model builds a text graph with texts as nodes and edges learned from current text representations via manifold learning and message passing. To further use unlabeled texts to improve few-shot performance, a knowledge distillation is devised to optimize the problem. This offers a trade-off between expressiveness and complexity. Experiments on social media datasets demonstrate the state-of-the-art performance of our model on attribute inferences with considerably fewer labeled texts.

Machine learning models have been widely used for fraud detection, while developing and maintaining these models often suffers from significant limitations in terms of training data scarcity and constrained resources. To address these issues, in this paper, we leverage machine learning vulnerability to adversarial attacks, and design a novel model AdvRFD that Adversarially Reprograms an ImageNet classification neural network for Fraud Detection task. AdvRFD first embeds transaction features into a host image to construct new ImageNet data, and then learns a universal perturbation to be added to all inputs, such that the outputs of the pretrained model can be accordingly mapped to the final detection decisions for all transactions. Extensive experiments on two transaction datasets made over Ethereum and credit cards have demonstrated that AdvRFD is effective to detect fraud using limited data and resources.

Machine learning, especially deep learning, has emerged as one of the most powerful tools for attribute inference attacks over social media, which poses serious threats to users' privacy and security. In this paper, we explore a novel perspective of protecting data privacy in social media, where we take advantage of the vulnerability of machine learning, and introduce adversarial attacks to forge latent feature representations and mislead attribute inference attacks. Considering that text data in social media shares the most significant privacy of users, we investigate how text-space adversarial attacks can be elaborated to obfuscate users' attributes, and accordingly present a text-space adversarial attack as defense, or AaaD for short. Specifically, we advance AaaD by constructing semantically and visually similar word candidates to perturb, and leveraging word importance scores as selection probabilities to upgrade a population-based optimization to expedite adversarial text generation. We evaluate the performance of AaaD on two social media data sets, while the experimental results validate its effectiveness against inference attacks. Our work yields great value and unveils a new insight on the applicability of adversarial attacks for attribute obfuscation and privacy protection.

Despite the apparent benefits of modern social coding paradigm such as Stack Overflow, its potential security risks have been largely overlooked (e.g., insecure codes could be easily embedded and distributed). To address this imminent issue, in this paper, we bring a significant insight to leverage both social coding properties and code content for automatic detection of insecure code snippets in Stack Overflow. To determine if the given code snippets are insecure, we not only analyze the code content, but also utilize various kinds of relations among users, badges, questions, answers and code snippets in Stack Overflow. To model the rich semantic relationships, we first introduce a structured heterogeneous information network (HIN) for representation and then use meta-path based approach to incorporate higher-level semantics to build up relatedness over code snippets. Later, we propose a novel hierarchical attention-based sequence learning model named CodeHin2Vec to seamlessly integrate node (i.e., code snippet) content with HIN-based relations for representation learning. After that, a classifier is built for insecure code snippet detection. Integrating our proposed method, an intelligent system named iTrustSO is accordingly developed to address the code security issues in modern software coding platforms. Comprehensive experiments on the data collections from Stack Overflow are conducted to validate the effectiveness of our developed system iTrustSO by comparisons with alternative methods.

To combat the evolving Android malware attacks, systems using machine learning techniques have been successfully deployed for Android malware detection. In these systems, based on different feature representations, various kinds of classifiers are constructed to detect Android malware. Unfortunately, as classifiers become more widely deployed, the incentive for defeating them increases. In this paper, we first extract a set of features from the Android applications (apps) and represent them as binary feature vectors; with these inputs, we then explore the security of a generic learning-based classifier for Android malware detection in the presence of adversaries. To harden the evasion, we first present count featurization to transform the binary feature space into continuous probabilities encoding the distribution in each class (either benign or malicious). To improve the system security while not compromising the detection accuracy, we further introduce softmax function with adversarial parameter to find the best trade-off between security and accuracy for the classifier. Accordingly, we develop a system named DroidEye which integrates our proposed method for Android malware detection. Comprehensive experiments on the real sample collection from Comodo Cloud Security Center are conducted to validate the effectiveness of DroidEye against adversarial Android malware attacks. Our proposed secure-learning paradigm is also applicable for other detection tasks, such as spammer detection in social media.

Because of the explosive growth of Android malware and due to the severity of its damages, the detection of Android malware has become an increasing important topic in cybersecurity. Currently, the major defense against Android malware is commercial mobile security products which mainly use signature-based method for detection. However, attackers can easily devise methods, such as obfuscation and repackaging, to evade the detection, which calls for new defensive techniques that are harder to evade. In this paper, resting on the analysis of Application Programming Interface (API) calls extracted from the smali files, we further categorize the API calls which belong to the some method in the smali code into a block. Based on the generated API call blocks, we then explore deep neural networks (i.e., Deep Belief Network (DBN) and Stacked AutoEncoders (SAEs)) for newly unknown Android malware detection. Using a real sample collection from Comodo Cloud Security Center, a comprehensive experimental study is performed to compare various malware detection approaches. The experimental results demonstrate that (1) our proposed feature extraction method (i.e., using API call blocks) outperforms using API calls directly in Android malware detection; (2) DBN works better than SAEs in this application; and (3) the detection performance of deep neural networks is better than shallow learning architectures.