Yong Xie

Large language models (LLMs) have demonstrated impressive abilities in various natural language processing tasks. While the LLMs’ hallucinations and the lack of up-to-date knowledge undermine their reliability in knowledge-intensive applications. Knowledge graphs can mitigate it by their structured and faithful source of factual knowledge, yet effectively integrating knowledge graphs into LLMs remains challenging due to issues in retrieval efficiency and structural understanding. To explore corresponding solutions, this paper proposes a novel Chain-of-Thought (CoT) enhanced path planning and retrieval (CoTPR) framework that combines Chain-of-Thought reasoning with path planning for knowledge graph question answering. Firstly, a general prompt format is designed to generate CoT sequences and relation paths from questions. Then, a two-stage searching method is proposed: (1) Based on the relation paths, a breadth-first search is utilized to obtain a knowledge subgraph. (2) The knowledge subgraph optimizes the query using CoT to enhance potential intermediate knowledge gradually. Extensive experiments are conducted on two benchmark datasets. The results show that CoTPR achieves state-of-the-art performance, with significant gains in Hits@1 and F1 scores. CoTPR paves the way for more reliable and efficient LLM agent-KG integration in complex reasoning tasks, advancing the autonomous capability of LLMs in structured knowledge environments.

Website Fingerprinting (WF) attacks based on Deep Neural Networks (DNN) pose serious threats to user privacy in anonymous communication systems such as Tor. Existing defenses show limited effectiveness in black-box scenarios, facing challenges including poor perturbation transferability, lack of directional and real-time constraints, limited universality, and high overhead. To address these issues, this paper proposes HAAD (Heuristic Ant-based Adversarial Defense), a black-box defense framework based on an ant colony algorithm. HAAD generates perturbations by combining an ensemble substitute model with ant colony optimization, avoiding reliance on gradient information and enhancing transferability to unknown models. Additionally, a per-packet injection method with directional constraints is designed to achieve real-time defense without extra latency. A universal perturbation algorithm is introduced, optimizing perturbations for diverse traffic scenarios via a scoring mechanism, effectively reducing communication overhead. Experimental results show that under various adversarial attack models, HAAD achieves a Defense Success Rate (DSR) exceeding 90% while requiring only 30 injected dummy packets—incurring minimal overhead. When deployed on a P4 switch, the maximum end-to-end latency is only 200 nanoseconds. We further implement HAAD in the Tor pt-plugin, where the full-page load median latency is merely 0.02 seconds. These results highlight HAAD's strong defensive capability and practical applicability.

Hashing retrieval plays a crucial role in image retrieval tasks due to its efficiency and accuracy. In singlemodal hashing-based image retrieval, the core challenge lies in enabling the network to fully understand the image content, thereby promoting more effective feature extraction. Inspired by cross-modal learning paradigms, we observe that augmenting images with captions or descriptive text provides rich contextual information that enhances the network's understanding of visual content. Incorporating textual descriptions into image representations helps overcome the limitations of relying solely on visual features, such as incomplete information or interpretive bias, which leads to more robust feature extraction. However, unresolved problems remain, including how to obtain and utilize text semantics when the dataset contains only images and how to effectively partition images to explore global structures without supervision. In this paper, we propose a novel Text-Guided Unsupervised Hashing (TGUH) method with community exploration. TGUH incorporates textual semantic representations into singlemodal image retrieval through multimodal information fusion, facilitating a deeper understanding of image content within the network. Additionally, we design an unsupervised dynamic community partitioning approach to adaptively partition image communities. This allows for the construction of hash centers under unsupervised conditions. Experimental results demonstrate that our method achieves significant improvements in retrieval accuracy, highlighting its practical advantages and potential for real-world applications.

Federated learning facilitates collaborative model training across multiple clients while preserving data privacy. However, its performance is often constrained by limited communication resources, particularly in systems supporting a large number of clients. To address this challenge, integrating over-the-air computations into the training process has emerged as a promising solution to alleviate communication bottlenecks. The system significantly increases the number of clients it can support in each communication round by transmitting intermediate parameters via analog signals rather than digital ones. This improvement, however, comes at the cost of channel-induced distortions, such as fading and noise, which affect the aggregated global parameters. To elucidate these effects, this paper develops a theoretical framework to analyze the performance of over-the-air federated learning in large-scale client scenarios. Our analysis reveals three key advantages of scaling up the number of participating clients: (1) Enhanced Privacy: The mutual information between a client’s local gradient and the server’s aggregated gradient diminishes, effectively reducing privacy leakage. (2) Mitigation of Channel Fading: The channel hardening effect eliminates the impact of small-scale fading in the noisy global gradient. (3) Improved Convergence: Reduced thermal noise and gradient estimation errors benefit the convergence rate. These findings solidify over-the-air model training as a viable approach for federated learning in networks with a large number of clients. The theoretical insights are further substantiated through extensive experimental evaluations.