Personal Knowledge Graphs (PKGs) organize an individual user's information into a structured format comprising entities, attributes, and relationships. By leveraging this structured and semantically rich data, PKGs have become essential for securing personal data management and delivering personalized services. To unlock their potential in personalized recommendations, prior research has explored the construction of PKGs and recommendation methods built upon them. However, these studies often overlook challenges associated with distributed PKGs across different users, such as joint training and privacy protection. To address these challenges, we propose PKGRec, a federated graph recommendation method specifically designed for PKGs, which utilizes a federated learning framework to ensure user privacy and data security during joint learning. Furthermore, to accommodate the user-centric graph structure of PKGs, our approach categorizes entities into three types: users, items, and other entities. It then applies a novel staged graph convolution method to model various entities based on these entity categories during local training. To enable efficient graph information sharing among distributed PKGs without requiring additional data transfer or aggregation, PKGRec performs graph expansion on the trained gradients by federated aggregation. Extensive experiments conducted on four publicly available datasets demonstrate that our method consistently outperforms the existing federated recommendation approaches.

We release a large-scale dataset that captures interactions between human users and CommentRobert, an LLM-based social media agent on Weibo. The dataset contains Weibo posts in which users actively mention the LLM agent account @CommentRobert, indicating that the users are interested in interacting with the platform-empowered LLM agent. The dataset contains 557,645 interactions from 304,400 unique users over 17 months. We detail our data collection methodology, user attributes, and content characteristics, underscoring the dataset's value in examining real-world human-LLM agent interactions. Our analysis offers insights into the demographic and behavioral traits of users interested in the selected LLM agent, interaction dynamics between humans and the agent, and linguistic patterns in comments. These interactions provide a unique lens through which to explore how humans perceive, trust, and communicate with LLMs. This dataset enables further research into modeling human intent understanding, improving LLM agent design, and studying the evolution of human-LLM agent relationships. Potential applications also include long-term user engagement prediction and AI-generated comment detection on social platforms. This constructed dataset is available at https://zenodo.org/records/16921462.

The data stream generated by users on web applications is often collected using a local differential privacy (LDP) approach to ensure privacy. This approach offers rigorous theoretical guarantees and low computational overhead, albeit at the expense of data utility. Data utility encompasses both the value of individual data points and the temporal relevance that exists between them, but existing studies primarily focus on enhancing the former utility while neglecting the latter. Furthermore, the collected data often requires cleaning, and we have demonstrated through a case study that data stream lacking time relevance poses a significant risk to users' privacy during the cleaning process. In this paper, for the first time we present an online LDP publishing mechanism while preserving the inherent temporal relevance for the infinite stream, called the Sampling Period Perturbation Algorithm (SPPA). Specifically, we model the temporal relevance between data points as the Fourier interpolation function, resulting in a computational complexity reduction from O(n2) to O(n log n) when compared with the conventional Markov approach in the offline setting. To strike a better balance between privacy and utility, we add noise to the sampling period due to its minimal impact on sensitivity, which is analyzed by our novel concepts of (ε,τ)-temporal indistinguishability and (ε,w,τ)-event LDP. Through extensive experiments, SPPA exhibits superior performance in terms of both data utility and privacy preservation compared to the state-of-the-art baselines. In particular, when ε=1, compared with the state-of-the-art baseline, SPPA diminishes the MSE by up to 64.2%, and raises the event monitoring efficiency by up to 21.4%.

Federated learning (FL) is a distributed machine learning paradigm in which clients collaboratively train models in a privacy-preserving manner. While centralized FL (CFL) suffers from single points of failure and performance bottlenecks, decentralized FL (DFL), which depends on inter-client communication, has emerged to eliminate the need of a central entity. However, due to lack of the coordination of a central server, heterogeneous data distribution across clients makes local models in DFL inclined to diverge towards their local objectives, resulting in poor model accuracy. Moreover, each client in DFL needs to communicate with multiple neighbors, yielding a heavy communication load. To tackle these challenges, we propose a novel DFL framework called DFLStar, which can improve DFL from two perspectives. First, to avoid significantly diverging towards local data, DFLStar incorporates self-knowledge distillation to enhance the local model training by assimilating knowledge from the aggregated model. Second, clients in DFLStar identify and only select the most informative neighbors (based on the last layer model similarity) for parameter exchange, thereby minimizing the communication overhead. Our experimental results on two real datasets demonstrate that DFLStar significantly improves both communication overhead and training time compared to traditional DFL algorithms while achieving a specific target accuracy. Furthermore, within a fixed training duration, DFLStar constantly obtains the highest model accuracy compared to the baselines.

Federated graph learning (FGL) has risen as a promising paradigm for collaboratively training graph neural networks while safeguarding data privacy. Nevertheless, the distributed nature of FGL also renders it susceptible to backdoor attacks. Although backdoor attacks are recognized as a significant threat to both centralized graph learning and federated learning (FL), the study of such attacks in FGL remains very limited. Current research on FGL backdoor attacks often merely adapts centralized graph backdoor attacks or FL backdoor attacks designed for image classification tasks to the FGL context, leaving key issues such as the effectiveness of triggers and the stealthiness of malicious models largely unexplored. To bridge this research gap, in this paper, we propose a novel backdoor attack, named FGLBA, targeting the FGL paradigm. Specifically, we design an input-aware trigger generator that generates a customized trigger for each target node based on its feature vector and neighborhood information, making that poisoned nodes injected with triggers are more likely misclassified into the category specified by the attacker. Additionally, we develop a stealthy federated backdoor training strategy that leverages collaborative optimization among multiple malicious clients to circumvent existing server-side defenses. The trigger generator and malicious clients' local models are iteratively optimized through a bilevel optimization framework, enabling the malicious models to achieve optimal attack performance under the optimal trigger generator. Extensive experiments on 4 real-world datasets demonstrate the effectiveness and superiority of our attack, outperforming all baseline attacks and successfully bypass 6 state-of-the-art and classical FL backdoor defenses.

Federated learning (FL) is an emerging paradigm using a parameter server (PS) to coordinate multiple decentralized clients for training a common model without exposing their raw data. Despite its amazing capability in preserving data privacy, FL confronts two significant challenges that have not been sufficiently addressed by existing works, which are: 1) heterogeneous data distributed on clients given that the PS cannot alter data locations. 2) limited computation resources of FL clients who may conduct model training with mobile devices. To tackle these challenges, we propose a novel federated one-shot transferring of intermediate features (FedOTF) algorithm. More specifically, FedOTF consists of two stages: feature extraction and model reconstruction. To overcome challenge 1, clients in stage 1 only collaboratively train a small model in a federated manner, with the objective to extract features. In stage 2, intermediate features (generated by the small model trained by stage 1) are transferred from clients to the PS so that a large model can be trained to overcome challenge 2. In addition, the design of FedOTF is robust, which can flexibly diminish the number of communication rounds of stage 1 when network capacity is limited, and reduce the amount of exposed features when privacy is concerned. To verify the superiority of FedOTF, we conduct comprehensive experiments with real datasets. The experiment results demonstrate that FedOTF can significantly improve the model utility of FL because a better model can be finally obtained by the PS without incurring heavy computational load on clients. Besides, we conduct robustness evaluation of FedOTF which can achieve stable performance when varying network capacity and privacy requirement.