Data heterogeneity is one of the key challenges in federated learning, and many efforts have been devoted to tackling this problem. However, distributed concept drift with data heterogeneity, where clients may additionally experience different concept drifts, is a largely unexplored area. In this work, we focus on real drift, where the conditional distribution $P(\mathcal{Y}|\mathcal{X})$ changes. We first study how distributed concept drift affects the model training and find that local classifier plays a critical role in drift adaptation. Moreover, to address data heterogeneity, we study the feature alignment under distributed concept drift, and find two factors that are crucial for feature alignment: the conditional distribution $P(\mathcal{Y}|\mathcal{X})$ and the degree of data heterogeneity. Motivated by the above findings, we propose FedCCFA, a federated learning framework with classifier clustering and feature alignment. To enhance collaboration under distributed concept drift, FedCCFA clusters local classifiers at class-level and generates clustered feature anchors according to the clustering results. Assisted by these anchors, FedCCFA adaptively aligns clients' feature spaces based on the entropy of label distribution $P(\mathcal{Y})$, alleviating the inconsistency in feature space. Our results demonstrate that FedCCFA significantly outperforms existing methods under various concept drift settings. Code is available at https://github.com/Chen-Junbao/FedCCFA.

APT attacks have posed serious threats to the security of cyberspace nowadays which are usually tailored for specific targets. Identification and understanding of APT attacks remains a key issue for society. Attackers often utilize malware as the weapons to launch cyber-attacks. For this reason, detecting APT malware and gaining an insight of its malicious behaviors can strengthen the power to understand and counteract APT attacks. Based on the above motivation, this paper proposes a novel APT malware detection and cognition framework named APTMalInsight aiming at identifying and cognizing APT malware by leveraging system call information and ontology knowledge. We systematically study APT malware and extracts dynamic system call information to describe its behavioral characteristics. With respect to the established feature vectors, the APT malware can be detected and clustered into their belonging families accurately. Furthermore, a horizontal comparison between APT malware and the traditional malware is conducted from the perspective of behavior types, to understand the behavioral characteristics of APT malware in depth. On the above basis, the ontology model is introduced to construct the APT malware knowledge framework to represent its typical malicious behaviors, thereby implementing the systematic cognition of APT malware and providing contextual understanding of APT attacks. The evaluation results based on real APT malware samples demonstrate that the detection and clustering accuracy can reach up to 99.28% and 98.85% respectively. In addition, APTMalInsight supplies an effective cognition framework for APT malware and enhances the capability to understand APT attacks.

With the rapid evolution of the Internet, the application of artificial intelligence fields is more and more extensive, and the era of AI has come. At the same time, adversarial attacks in the AI field are also frequent. Therefore, the research into adversarial attack security is extremely urgent. An increasing number of researchers are working in this field. We provide a comprehensive review of the theories and methods that enable researchers to enter the field of adversarial attack. This article is according to the “Why? → What? → How?” research line for elaboration. Firstly, we explain the significance of adversarial attack. Then, we introduce the concepts, types, and hazards of adversarial attack. Finally, we review the typical attack algorithms and defense techniques in each application area. Facing the increasingly complex neural network model, this paper focuses on the fields of image, text, and malicious code and focuses on the adversarial attack classifications and methods of these three data types, so that researchers can quickly find their own type of study. At the end of this review, we also raised some discussions and open issues and compared them with other similar reviews.

It is a wide-spread way to detect malware by analyzing its behavioral characteristics based on API call sequences. However, previous studies usually just focus on its static or dynamic API call sequence, while neglecting the correlation between them. Our experimental results show that there exists an underlying relation between the dynamic and static API call sequences of malware. The relation can be described as “the syntax is different, but the semantics is similar”. Based on this discovery, this paper first attempts to explore the difference and relation between the static and dynamic API sequences of malicious programs. We correlate and fuse their dynamic and static API sequences into one hybrid sequence based on semantics mapping and then construct the hybrid feature vector space. Furthermore, we mine and define the malicious behavior types of the programs, and provide explainable results for malware detection. Our study has addressed the shortcoming of the previous approaches that they usually pay attention to detection but neglect explanation. By correlation and fusion of the static and dynamic API sequences, we establish an explainable malware detection framework, called MalDAE. The evaluation results show that the detection and classification accuracy of MalDAE can reach up to 97.89% and 94.39% respectively outperforming the previous similar studies by comprehensive comparison. In addition, MalDAE gives an understandable explanation for common types of malware and provides predictive support for understanding and resisting malware.

To handle the security threat faced by the widespread use of Internet of Things (IoT) devices due to the ever-lasting increase of malware, the security researchers increasingly rely on machine learning techniques based on various static and/or dynamic features. Unfortunately, the state-of-the-art detection techniques may fail to identify the malware effectively because the malware is often obfuscated to camouflage its characteristics and thwart the analysis process. In order to identify the disguised malware accurately, a malware detection framework named MalInsight is proposed by profiling malware from three aspects which are basic structure, low-level behavior, and high-level behavior. These aspects reflect the structural features, the underlying operations interacting with the OS, and the operations on the files, the registry, and the network respectively. Based on the above findings, an accurate and rich feature space is built which enables to depict and detect malware more effectively. In order to validate the effectiveness of MalInsight, an extensive experiment is conducted on a real-world malware dataset. Our experimental results show that MalInsight can detect not only obfuscated malware instances with an accuracy of 99.76% but also unseen and new malware with an accuracy of 97.21%. Furthermore, MalInsight can classify the malware samples into their families with an accuracy of 94.2% outperforming the typical detection approach based on the API sequence as the dynamic behavior features by almost 9%. In addition, the importance of the three aspects is evaluated and sorted quantitatively demonstrating that these aspects play the same effects with the optimal feature set.