To analyze the massive social networks for providing personalized services, community search is widely studied to find the densely connected subgraph that can reflect the network properties for a given query. The existing community search methods adopt single community model to make structural constraints on communities, which can only describe single interaction mode. Since they fail to capture the semantics of the network with multiple interaction modes, they struggle to find the representative communities. To solve this issue, we design a novel community model called ( τ, ρ )-camp to flexibly capture complex network semantics in any level of granularity. We propose the unified support maximized community search problem to find the communities with the densest network semantics, which is proven a NP-hard problem. By constructing a hierarchical index structure, we propose an approximate community search algorithm with approximation ratio of 2 and linear time complexity of the query size. Extensive experiments are conducted on two public datasets and two crawled datasets. The experimental results prove the effectiveness and efficiency of our method.

With the deployment of the QUIC protocol, website fingerprinting attacks targeting QUIC traffic are becoming a growing concern. Since the deployment is incremental, attackers must continuously crawl the QUIC traffic of new QUIC-enabled websites to update their attack models. For the latency caused by data crawling and classifier training, existing few-shot website fingerprinting (FSWF) attacks rely on representation learning to mitigate data dependency. To further achieve zero-latency identification, TCP traffic can be applied to construct the attack model before QUIC deployment. However, the different protocol semantics of TCP and QUIC lead to differences in the latent features. As representation learning models cannot eliminate the website feature differences, classifiers trained on TCP-based features are difficult to adapt to QUIC traffic. To address the issue, we propose a novel cross-protocol FSWF attack method to fuse cross-protocol website features. The proposed method forces TCP features and QUIC features to be in the same feature space by sharing model parameters, and reduces cross-protocol website feature differences through inter-protocol adversarial representation learning. Meanwhile, it utilizes a non-linear classifier to fit the fused features. The proposed method enables zero-latency identification for QUIC traffic based on a few TCP traffic. We conducted comprehensive evaluation experiments on public datasets from both closed-world and open-world settings. The proposed method outperforms state-of-the-art methods in zero-latency identification.

Few-shot website fingerprinting (FSWF) attacks attempt to identify whether the users have access to specific websites based on a few training data. Existing FSWF attack methods focus on adapting to variable network conditions in real scenarios. They use various techniques to transfer the model to adapt to test data which has a different distribution from training data. However, recent methods ignore the impact of pre-training data diversity on adaptability. The poor data diversity caused by the user-specific data crawl limits representation ability, and further hinders rapid adaptation to new network conditions. Due to the extreme Non-IId between multiple attackers’ datasets, it is not feasible to mix multiple datasets or perform traditional federated learning methods to improve representation ability. To address the issue, we propose a novel method based on a joint learning framework to achieve the collusion FSWF attacks. The proposed method fuses the feature spaces of multiple user-side attackers to enhance the representation ability of the local model, and constructs a virtual fusion center to mitigate the impact of Non-IID. It improves the adaptability under variable network conditions for the local attacker. This paper conducts comprehensive experiments to evaluate the performance of the proposed method in both closed-world and open-world settings. Compared with the state-of-the-art method, the proposed method improves the accuracy by up to 13.02% in the closed-world setting and the AUC by up to 0.085 in the open-world setting, respectively.

Anchor link prediction enhances the effectiveness of digital forensics through the identification of multiple social network users. The current methods based on deep learning are characterized by both the exaggerated similarity between adjacent nodes in the same latent space and the variation in the feature spaces caused by semantics. A novel approach is developed to fuse the semantic features of different networks in this paper. The proposed method is divided into two stages. Firstly, representation learning pays more attention to the influence of uncertainty on the equivalence of node network structure, and introduces the difference between adjacent nodes from the latent space. Secondly, a joint representation learning framework trains and exchanges the parameters depending on known anchor links. The joint representation learning framework injects fused features into the representation learning processes of different networks. The combination of enhanced discrimination and cross-network feature fusion reduces the feature space differences caused by the semantics of different social networks. This paper conducts comprehensive experiments on social networks in the real world. The outcome shows that the proposed approach is more efficient and robust compared to the existing state-of-theart methods.

Anchor link prediction exacerbates the risk of privacy leakage via the de-anonymization in edge computing. The predictive effect of traditional unsupervised learning methods is too dependent on user attributes and supervised learning methods are sensitive to network structure noise. Anchor link prediction methods based on graph embedding are restricted by the sparsity of the observable anchor links which can be used for training. To facilitate the effectiveness and robustness of the anchor link prediction, we have proposed a novel method which reduces the restrictions on the observable anchor links used for training. The proposed method consists of two phases. First, graph embedding based on network structural roles is used to generate the latent feature space, reconciling the distinction and similarity between nodes. Second, the supervised learning for optimizing the Wasserstein distance which estimates the minimum amount of work to change one distribution into the other. The combination of the reconciled latent feature space and the estimate for the amount of change alleviates the restriction on observable overlapping parts. Extensive experiments on real-life social networks have demonstrated that the proposed method significantly outperforms the state-of-the-art methods in terms of both precision and robustness.

Short hairpin RNA (shRNA)-mediated gene silencing is an important technology to achieve RNA interference, in which the design of potent and reliable shRNA molecules plays a crucial role. However, efficient shRNA target selection through biological technology is expensive and time consuming. Hence, it is crucial to develop a more precise and efficient computational method to design potent and reliable shRNA molecules. In this work, we present an interpretable classification model for the shRNA target prediction using the Light Gradient Boosting Machine algorithm called ILGBMSH. Rather than utilizing only the shRNA sequence feature, we extracted 554 biological and deep learning features, which were not considered in previous shRNA prediction research. We evaluated the performance of our model compared with the state-of-the-art shRNA target prediction models. Besides, we investigated the feature explanation from the model's parameters and interpretable method called Shapley Additive Explanations, which provided us with biological insights from the model. We used independent shRNA experiment data from other resources to prove the predictive ability and robustness of our model. Finally, we used our model to design the miR30-shRNA sequences and conducted a gene knockdown experiment. The experimental result was perfectly in correspondence with our expectation with a Pearson's coefficient correlation of 0.985. In summary, the ILGBMSH model can achieve state-of-the-art shRNA prediction performance and give biological insights from the machine learning model parameters.