With the development and popularization of artificial intelligence technologies represented by deep learning, the security issues they continuously expose have become a huge challenge affecting cyberspace security. Traditional cloud-centric distributed machine learning, which trains models or optimizes model performance by collecting data from participating parties, is susceptible to security attacks and privacy attacks during the data exchange process, leading to consequences such as a decline in overall system efficiency or the leakage of private data. Federated learning, as a distributed machine learning paradigm with privacy protection capabilities, exchanges model parameters through frequent communication between clients and parameter servers, training a joint model without the raw data leaving the local area. This greatly reduces the risk of private data leakage and ensures data security to a certain extent. However, as deep learning models become larger and federated learning tasks more complex, communication overhead also increases, eventually becoming a barrier to the application of federated learning. Therefore, the exploration of communication optimization methods for federated learning has become a hot topic. The technical background and workflow of federated learning were introduced, and the sources and impacts of its communication bottlenecks were analyzed. Then, based on the factors affecting communication efficiency, existing federated learning communication optimization methods were comprehensively sorted out and analyzed from optimization objectives such as model parameter compression, model update strategies, system architecture, and communication protocols. The development trend of this research field was also presented. Finally, the problems faced by existing federated learning communication optimization methods were summarized, and future development trends and research directions were looked forward to.

In the data era, threats to personal privacy information in ubiquitous sharing environments are widespread, such as apps frequently collecting personal information beyond scope, and big data-enabled price discrimination against frequent customers. The need for multi-party privacy computing for cross-system exchanges is urgent. This work focused on the needs of multi-party privacy computing for cross-system exchanges in ubiquitous sharing environments, taking the security sharing and controlled dissemination of private data in multi-party data fusion applications as the starting point, and provided reviews of existing relevant work from the perspectives of multi-party privacy computing, multi-party privacy information sharing control, and multi-party data collaborative secure computing. First, the background and research status of personal privacy information protection in a ubiquitous sharing environment were analyzed. Then, the latest domestic and foreign research results in recent years regarding multi-party privacy computing, multi-party privacy information sharing control, and multi-party data collaborative security computing were reviewed and comparatively analyzed. Regarding multi-party privacy computing, technologies such as full lifecycle privacy protection, privacy information flow control, and secure exchange of sensitive data were introduced. In terms of multi-party privacy information sharing control, localized control, extended control, and anonymization control techniques were discussed. In the aspect of multi-party data collaborative secure computing, commonly used techniques in both academia and industry were discussed. Finally, the challenges and development directions of multi-party privacy computing were prospected. There were still limitations for anonymity, scrambling, or access control-based traditional privacy desensitization measures, cryptography-based measures, and federated learning-based measures, while privacy computing theory provided a computational and information system framework for full-lifecycle protection, which needed to be combined with different application scenarios to implement full-lifecycle privacy information protection.

To improve the performance of coverage-guided fuzzing, a method for self-adaptive optimization of fuzzing using distribution divergence and a deep reinforcement learning model was proposed. An interprocedural comparison flow graph was first constructed based on the interprocedural control flow graph to characterize the spatial random field corresponding to the branch condition variables of the program under test, and the distribution features of the random field generated by a fuzzing mutation strategy were extracted using the Monte Carlo method. Then, a deep graph convolutional neural network was constructed to extract the feature embeddings of the interprocedural comparison flow graph, and this neural network was used as the deep Q-network for deep reinforcement learning. Finally, an online deep reinforcement learning model was established based on the dual deep Q-network model, and an intelligent agent was trained to optimize the fuzzing mutation strategy. This deep reinforcement learning model defined its state using the random field distribution features corresponding to the seed file and the associated blocks. The selection for the focused mutation block of a seed file was defined as an action, and the distribution divergence of the approximate distributions of the random fields before and after the action was defined as the reward. A prototype was implemented for this fuzzing optimization method, and multiple rounds of up to 24 hours of evaluation were carried out on this prototype. The experimental results show that on the benchmark FuzzBench, the code coverage speed and overall coverage achieved by the prototype are significantly better than those of the baseline fuzzer AFL++ and HavocMAB, and better results are achieved on most benchmarks compared to CmpLog. On the benchmark Magma, stronger vulnerability triggering capability is demonstrated by the prototype on the benchmarks openssl, libxml, and sqlite3.

Deep neural networks have become an integral part of people's daily lives. However, researchers observed that these networks were susceptible to threats from adversarial samples, leading to abnormal behaviors such as misclassification by the network model. The presence of adversarial samples poses a significant threat to the application of deep neural networks, especially in security-sensitive scenarios like license plate recognition systems. Presently, most existing defense and detection technologies against adversarial samples show promising results for specific types of adversarial attacks. However, they often lack generality in addressing all types of adversarial attacks. In response to adversarial sample attacks on real-world license plate recognition systems, an unsupervised adversarial sample detection system named FIAD was proposed, which was based on analyzing the inherent variations in neural networks trained on clean samples and the dimensional complexity between clean samples. FIAD utilized neural network invariants and local intrinsic dimensionality invariants for effective sample detection. The detection system was deployed into widely used open-source license plate recognition systems, HyperLPR and EasyPR, and extensive experiments were conducted using the real dataset CCPD. The results from experiments involving 11 different types of attacks indicate that, compared to 4 other advanced detection methods, FIAD can effectively detect all these attacks at a lower false positive rate, with an accuracy consistently reaching 99%. Therefore, FIAD exhibits good generality against various types of adversarial attacks.

Federated learning, as a distributed machine learning architecture, can complete model co-training while protecting data privacy, and is widely used in Artificial Intelligence of Things. However, there are often security threats such as privacy breaches and poisoning attacks in federated learning. In order to overcome the performance and security challenges of using federated learning for joint training among multiple institutions in the context of intelligent Internet of Things, a two-level federated security learning architecture was proposed for intelligent Internet of Things. The entire security learning system was divided into a two-level architecture of bottom and top layers. The bottom architecture consisted of various IoT devices and a server within the organization. Different devices were connected through blockchain networks, and the server detected and eliminated malicious devices through the historical gradient uploaded by the devices, avoiding slow convergence and decreased global model accuracy caused by poisoning attacks. The top-level architecture consisted of servers from different institutions, using secure multi-party computation based on secret sharing for secure aggregation, protecting gradient privacy while achieving decentralized gradient aggregation. The experimental results show that the architecture achieves detection accuracy of over 85% for four common poisoning attacks, greatly improving the security of the system and achieving decentralized security aggregation with gradient privacy protection within linear time complexity.

Attribute-based access control is fundamentally dependent on the secure service attribute calibration of object sources. Traditionally, the calibration of secure service attribute for textual data has been primarily reliant on human experts and machine learning methods, yet the efficiency and few-shot ability are insufficient. Moreover, traditional approaches have predominantly utilized entities in textual data as service attributes, resulting in coarse granularity, uncontrollable scale and management level, further leading to the problem of attribute-space explosion. Thus, a secure service attribute generation mechanism for textual data (IDSAM) was introduced. This mechanism addressed the aforementioned challenges by transforming the extraction of candidate service attributes, previously a sequence-calibrated problem, into a controllable-generation problem through instruction learning and in-context learning. Subsequently, WordNet was employed to achieve semantic deduplication and generalization of the candidate service attributes. Concurrently, to prevent semantic loss due to over-generalization, a cosine similarity threshold was regulated, enabling the generation of a service attribute set. Finally, a weighted directed acyclic attribute graph was constructed based on the similarity between initial and derived attributes within the set, facilitating the dynamic construction of a secure service attribute library with a controllable scale and adjustable security granularity, in accordance with security control requirements. The candidate service attribute extraction component of the proposed mechanism achieves an optimal average F1 score in few-shot experiments on the CoNLL-2003 dataset, surpassing the baseline model. This positions the mechanism as state-of-the-art. Furthermore, the mechanism is capable of dynamically mining secure service attributes with adjustable security control levels and controllable scales to meet varying security management requirements. The experimental results indicate that the proposed mechanism is effective in generating secure service attributes with the desired characteristics.

With the maturation of computer graphics (CG) technology in the field of image generation, the realism of created images has been improved significantly. Although these technologies are widely used in daily life and bring many conveniences, they also come with many security risks. If forged images generated using CG technology are maliciously used and widely spread on the Internet and social media, they may harm the rights of individuals and enterprises. Therefore, an innovative cross-stream attention enhanced central difference convolutional network was proposed, aiming at improving the accuracy of CG image detection. A dual-stream structure was constructed in the model, in order to extract semantic features and non-semantic residual texture features from the image. Vanilla convolutional layers in each stream were replaced by central difference convolutions, which allowed the model to simultaneously extract pixel intensity information and pixel gradient information from the image. Furthermore, by introducing a cross-stream attention enhancement module, the model enhanced feature extraction capability at the global level and promoted complementarity between the two feature streams. Experimental results demonstrate that this method outperforms existing methods. Additionally, a series of ablation experiments further verify the rationality of the proposed model design.

To comply with legal requirements for personal data privacy protection, mobile App developers typically disclose their data collection practices to users through privacy policies. Researchers have proposed various methods using natural language processing (NLP) techniques to analyze privacy policy texts and perform compliance checks. However, most existing studies focus on principles like transparency, openness, and legality, leaving a gap in the evaluation of the ‘minimum necessary’ principle. For this purpose, a framework called MNPD (minimum necessary principle detection) was proposed for automated compliance checking of applications from the perspective of the ‘minimum necessary’ principle. Initially, a multi-label text classification model categorized the target App based on its service type to determine the range of ‘minimum necessary information’ for different App categories. Then, prompt words were constructed to guide the large language model in extracting data collection practices of the App under its basic business functionality mode, transforming them into privacy statement triples and standardizing them. Finally, the compliance checking model conducted consistency checks on the text representation of the target App and evaluated its adherence to the ‘minimum necessary’ principle. The experimental results show that the proposed method achieves 86.20% F1 score in the automated analysis of 101 ‘Online Audio-Visual’ Apps obtained from Huawei’s application market.

Multi-key fully homomorphic encryption (MK-FHE) technology supports homomorphic operations on ciphertexts encrypted with different keys, and can be directly applied to real-world multi-user data fusion computing scenarios, such as multi-party collaborative computing and federated learning. At present, mainstream multi-key fully homomorphic encryption mainly uses bootstrapping technology to achieve multi-key homomorphic computing of LWE (learning with errors) encrypted ciphertexts. The time efficiency and storage overhead of expanding single-key ciphertexts to multi-key ciphertexts are closely related to the bootstrapping technology, which usually requires a large number of homomorphic evaluation keys and complex operations. Therefore, designing a multi-key fully homomorphic encryption scheme in combination with a better bootstrapping technology to improve computational efficiency and reduce storage overhead had become a key research issue. Based on the NTRU (number theory research unit) bootstrapping technology, an efficient multi-key fully homomorphic encryption scheme for LWE ciphertexts was designed. Compared with other schemes, the proposed scheme exhibited higher computational efficiency in the ciphertext expansion process. In addition, the correctness of the proposed scheme was elaborated and analyzed in detail. The proposed scheme was compared with the existing mainstream multi-key fully homomorphic encryption scheme in theory. The results show that the proposed scheme has better computational efficiency. Finally, the potential application of the scheme in cross-departmental supervision scenarios of multi-industry and multi-source data sales data was explored, which helped the tax department to realize tax verification under the premise of protecting data privacy, and helped promote the digital transformation and healthy development of various industries.

Website fingerprinting usually identifies the target website visited by users based on the website fingerprint characteristics exposed in the web traffic. It is essential in tracking users’ anonymous access behaviors and improving the anonymous traffic governance, especially on Tor network flows. However, many defense mechanisms emerged to disturb the distinctive traffic patterns, which results in website fingerprint identification failure. The existing website fingerprint identification method with the best robustness named RF can maintain good performance against various defense methods, but it is difficult to resist the targeted defense method RF Countermeasure. An anti-defense website fingerprinting based on hybrid feature matrix (ADF) was proposed. Unlike RF, ADF used the cumulative packet length instead of the cumulative packet number as the packet-level feature. On the basis of analyzing information leakage value of flow features, ADF constructed the robust flow features of the session level using packet direction distribution and the number of continuous packets in the same direction. Subsequently, a hybrid feature matrix (HFM) was constructed to resist various defense disturbance by combining the features of both packet-level and session-level. With the matrix as input, a robust flow classifier with convolutional neural network was established. Through extensive experimental analysis on the dataset provided by DF, the accuracy under RF Countermeasure is 95.4%, which is 21.2% higher than RF. This method also maintains good identification performance under other state-of-the-art defenses.

The personal information protection law (PIPL) of the People’s Republic of China served as an important legal framework for safeguarding personal information rights. It established clear regulations for personal information controllers in their activities involving the collecting, storing, using, and sharing of personal information. It also required that these controllers provide explanations within their privacy policies for the services they offered. This meant that any company providing services in China must first offer a privacy policy that complied with the requirements of the PIPL. Therefore, in order to analyze the compliance of privacy policies with respect to the PIPL, an intelligent method was presented for assessing privacy policy compliance based on a knowledge graph. First, a comprehensive analysis of the PIPL was conducted, and a multi-level privacy policy knowledge graph was proposed that covered concepts related to information protection that needed to be explained in privacy policies. Next, a semi-automated method was built for collecting privacy policies and collected the privacy policies of 400 Chinese Apps. 100 policies were cross-annotated based on the knowledge graph, resulting in the creation of the first Chinese privacy policy corpus tailored to the PIPL called APPCP-100 (APP-privacy-policy-corpus-for-PIPL-100). Using this corpus, a Chinese concept classifier model CPP-BERT was constructed to achieve efficient detection of privacy policy compliance. Finally, the knowledge graph was applied to conduct a comprehensive compliance analysis of privacy policies, and the results indicate that the current compliance of Chinese App privacy policies with the fine-grained concepts of the PIPL still needs improvement.

Smart internet of things (IoT), with its high flexibility, adaptability, and continuous evolution, brings both new challenges and enhanced user experiences. In particular, the endogenous security and secure transmission issues faced by IoT edge-area smart terminals, due to uncertain threats like unknown vulnerabilities and backdoors, are of considerable concern in the realm of smart IoT. To address these challenges, a security protection method based on mimic defense and Paillier encryption for smart IoT terminals was proposed. First, a security architecture was constructed based on the mimic defense theory to ensure the reliability and robustness of the system even when the edge domain smart terminal had its own security genetic defects. Then, a secure blind signature with Paillier encryption (SBSPE) algorithm was designed in this architecture to safeguard the data and privacy of the edge-area smart terminals throughout the data processing lifecycle of edge-area smart terminals. Blind signature technology was integrated into SBSPE algorithm on the basis of Paillier semi-homomorphic encryption algorithm. During data encryption, it employed blind factor technology to execute a blind signature, rendering ciphertext information indecipherable even if an attacker obtained access to the decryption key of the Paillier encryption algorithm. It also effectively supported the efficient and reliable transmission of data of the intelligent IoT devices. Finally, taking the mimic smart IoT system as an application scenario, the proposed method was theoretically analyzed and experimentally validated for its improved performance and security.