Yiwei Liu

Malicious URL attacks are becoming more domain-specific, for example, phishing tactics varying greatly between sectors, such as education and e-commerce. Therefore, cross-domain collaboration and knowledge sharing are essential for developing effective solutions to combat these threats. However, such collaboration is often hindered by privacy concerns and business sensitivities. Federated learning addresses these challenges by enabling multi-client collaboration without direct data exchange. Unfortunately, if highly expressive Transformer models are used, clients may face intolerable computational burdens, and the exchange of weights could quickly deplete network bandwidth. In this paper, we propose Fed-urlBERT, a federated learning method for malicious URL detection using pretrained Transformers. It leverages split learning to partition the Transformer encoder into client-side embeddings and server-side encoding. In this setup, only generate initial embeddings from local data, while the server handles the computationally intensive encoding. Our approach addresses privacy concerns and cross-domain collaboration needs in malicious URL defense while minimizing the client computational burden. We employ two aggregation strategies, FedAvg and FedALA, with the latter designed to mitigate client heterogeneity by integrating both local and global model information during each training round. We evaluate our approach across three distinct data distribution scenarios, including Independent and Identically Distributed (IID data) and two types of Non-IID data settings. The experimental results demonstrate that our method consistently achieves performance comparable to that of centralized models. Significantly, under balanced data conditions, our federated model also demonstrates a notable reduction in the False Positive Rate (FPR), highlighting a potential regularization benefit of the federated approach. The source code is accessible at https://github.com/Davidup1/FedURLBERT.

The field of biological research has been profoundly impacted by the emergence of biological pre-trained models, which have resulted in remarkable advancements in life sciences and medicine. However, the current landscape of biological pre-trained language models suffers from a shortcoming, i.e., their inability to grasp the intricacies of molecular interactions, such as ncRNA-protein interactions. It is in this context that our paper introduces a two-tower computational framework, termed RPIPLM, which brings forth a new paradigm for the prediction of ncRNA-protein interactions. The core of RPIPLM lies in its harnessing of the pre-trained RNA language model and protein language model to process ncRNA and protein sequences, thereby enabling the transfer of the general knowledge gained from self-supervised learning of vast data to ncRNA-protein interaction tasks. Additionally, to learn the intricate interaction patterns between RNA and protein embeddings across diverse scales, we employ a fusion of scaled dot-product self-attention mechanism and Multi-scale convolution operations on the output of the dual-tower architecture, effectively capturing both global and local information. Furthermore, we introduce supervised contrastive learning into the training of RPIPLM, enabling the model to effectively capture discriminative information by distinguishing between interacting and non-interacting samples in the learned representations. Through extensive experiments and an interpretability study, we demonstrate the effectiveness of RPIPLM and its superiority over other methods, establishing new state-of-the-art performance. RPIPLM is a powerful and scalable computational framework that holds the potential to unlock enormous insights from vast biological data, thereby accelerating the discovery of molecular interactions.

A large number of bug reports are submitted by users and developers in bug-tracking system every day. It is time-consuming for software maintainers to assign bug reports to appropriate developers for fixing manually. Many bug-triaging methods have been developed to automate this process. However, most previous studies mainly focused on analyzing textual content and failed to make full use of the structured information embedded in the bug-tracking system. In fact, this structured information, which plays an important role in bug triaging, reflects the process of bug tracking and the historical activities. To further improve the performance of automatic bug triaging, in this study, we propose a new representation learning model, PTITransE, for knowledge bases, which extends TransE via enhancing the embeddings with textual entity descriptions and is more suitable for bug triaging. Moreover, we make the first attempt to apply knowledge base and link prediction techniques to bug triaging. For each new bug report, the proposed framework can recommend top-k developers for fixing the bug report by using the learned embeddings of entities and relations. Evaluation is performed on three real-world projects, and the results indicate that our method outperforms baseline bug triaging approaches and can alleviate the cold-start problem in bug triaging.

The limited transaction throughput performance is the primary challenge for single-chain structure blockchain in practical applications. Directed acyclic graph (DAG) technology offers a novel topological structure, significantly improving blockchain’s ability to process massive transactions by containing numerous blocks in a parallel way. However, the existing parallel-chain schemes embed DAG into the Practical Byzantine Fault Tolerance (PBFT) protocol stages to expand the instance, which requires an extent of overall synchronization between multiple chains. Overall consensus synchronously among all parallel chains may delay the system’s progress, particularly when the graph structure is unevenly distributed. This paper introduces a new DAG-based blockchain named Fountain, which aims to provide a loosely coupled consensus algorithm based on the classic parallel-chain scheme Hashgraph. The Fountain scheme changes consensus finality from an overall behavior to a separate behavior, thereby reducing block commit latency from the perspective of each parallel chain. Theoretical analysis reveals that our scheme maintains the same security level as Hashgraph, while experimental simulations demonstrate its effectiveness in optimizing commit latency.

Deep Neural Networks (DNNs) have made significant progress in addressing the Non-Intrusive Load Monitoring (NILM) task, which aims to disaggregate appliance-level power signals from aggregated meter readings. Despite these advancements, existing DNN-based NILM approaches rely on training with power signals from individual appliances, which are obtained intrusively through sensor installations. This method is not only expensive but also poses a risk of damaging the original circuits. To overcome these limitations, we introduce the State-based Supervised NILM (SS-NILM) problem. Instead of using appliance power signal labels, we leverage on-off state information, which can be collected in a non-intrusive manner. However, solving SS-NILM presents a challenge, as it requires developing a model that maps on-off state labels to the corresponding appliance power signals in an indirectly supervised setting. In this work, we propose a state-based DNN that predicts the power signals of multiple target appliances simultaneously. The model is trained by minimizing the discrepancy between the aggregated prediction and the true aggregated power signal. Additionally, the model predicts the on-off states of appliances, which are used as auxiliary information to improve the accuracy of power signal predictions. Extensive experiments conducted on real-world datasets demonstrate that our model, trained using non-intrusive on-off state information, achieves performance comparable to that of traditional NILM models.

The increasing proliferation of GPS-enabled mobile devices, including Unmanned Aerial Vehicles (UAVs), smartphones, and laptops, has resulted in a significant upsurge in the outsourcing of spatial data to cloud servers for storage and computation purposes, such as task allocation and location-based services. However, the reliance on untrusted cloud servers introduces the risk of privacy breaches, as these servers possess the ability to deduce and access users’ private information based on task content and query requirements. Existing privacy-preserving task-allocation schemes offer only coarse-grained and non-temporal access control, which restricts their applicability in scenarios involving multiple users and time-series data, such as trajectory and time-related routes. To overcome these challenges, this paper proposes an Efficient and Privacy-Preserving Task Allocation with Temporal Access Control (EPTA-T) scheme for mobile crowdsensing. By leveraging the techniques of Gray code and randomizable matrix multiplication, EPTA-T achieves efficient and privacy-preserving task allocation in mobile crowdsensing. Specifically, EPTA-T supports fine-grained and temporal access control through the utilization of an attribute-based access tree and function integration. The formal security analysis demonstrated that EPTA-T effectively guarantees data privacy and query privacy throughout the task allocation process. Extensive experiments conducted using a real-world dataset indicated that the EPTA-T scheme surpassed the performance of the state-of-the-art scheme.

Phishing poses a significant threat to the financial and privacy security of internet users and often serves as the starting point for cyberattacks. Many machine-learning-based methods for detecting phishing websites rely on URL analysis, offering simplicity and efficiency. However, these approaches are not always effective due to the following reasons: (1) highly concealed phishing websites may employ tactics such as masquerading URL addresses to deceive machine learning models, and (2) phishing attackers frequently change their phishing website URLs to evade detection. In this study, we propose a robust, multi-view Transformer model with an expert-mixture mechanism for accurate phishing website detection utilizing website URLs, attributes, content, and behavioral information. Specifically, we first adapted a pretrained language model for URL representation learning by applying adversarial post-training learning in order to extract semantic information from URLs. Next, we captured the attribute, content, and behavioral features of the websites and encoded them as vectors, which, alongside the URL embeddings, constitute the website’s multi-view information. Subsequently, we introduced a mixture-of-experts mechanism into the Transformer network to learn knowledge from different views and adaptively fuse information from various views. The proposed method outperforms state-of-the-art approaches in evaluations of real phishing websites, demonstrating greater performance with less label dependency. Furthermore, we show the superior robustness and enhanced adaptability of the proposed method to unseen samples and data drift in more challenging experimental settings.