Vertical Federated Learning (VFL) enables collaborative modeling across different institutions without data leaving the local domains, and has been widely applied in highly sensitive scenarios such as financial risk control and medical diagnosis in recent years. However, the existing model-splitting architecture of VFL suffers from severe risks of label leakage. Malicious participants can exploit intermediate gradient feedback to launch label inference attacks, thereby recovering highly sensitive private labels and causing catastrophic privacy breaches. To address this issue, prior studies have proposed corresponding defense mechanisms, but they either require the introduction of additional models, resulting in excessive overhead, or severely degrade model performance. To strike a balance between model performance and privacy protection, this paper proposes LISS, a defense mechanism against label inference attacks. By dynamically disrupting the association between the gradient updates observed by the attacker and the ground-truth labels, this mechanism effectively blocks the attacker’s path to reconstructing a stable mapping model, thereby fundamentally thwarting label inference attacks. Furthermore, by combining lightweight pseudo-random operations, IPW-based batch-mean gradient correction, and residual shrinkage, it avoids the computational overhead of auxiliary models while retaining most of the main-task accuracy in the evaluated settings. Extensive experimental results demonstrate the superiority of our mechanism: it successfully reduces the success rates of various label inference attacks to near-random-guessing levels (e.g., dropping from 82.35
As artificial intelligence (AI) systems become embedded in high-stakes environments, security assessment must account for both the vulnerabilities of individual models and the dependencies through which their effects may influence a larger system. A central challenge is to connect these analyses with mitigation decisions while preserving the conditions and evidence that support each conclusion. This paper proposes a formal, lifecycle-aware security-testing framework that brings model-centric and system-centric analysis into a common assessment process. Within a finite scope, it relates threats to their data and deployment contexts, assesses mitigation outcomes using metric-specific criteria, and separates individual-defense evaluation from defense selection. To preserve the basis of these decisions, linked records associate findings with their source conditions and with any assumptions used in subsequent interpretation. A supporting reporting specification organizes this evidence for engineering and organizational review. Selected elements of the method are illustrated through reported digital-patch outcomes and dependency calculations in a perception example, with empirical observations distinguished from analytical assumptions.
Protecting sensitive information while maintaining data utility is a fundamental challenge in privacy-preserving data mining. Differential Privacy (DP)-based K-means clustering provides a principled solution by injecting noise through a privacy budget; however, determining an optimal budget allocation remains a complex task due to the inherent trade-offs among privacy protection, clustering utility, and model stability. Motivated by the effectiveness of Evolutionary Algorithms (EAs) in solving complex optimization problems, this paper proposes an enhanced Adaptive Multi-Objective Memetic Algorithm for Differentially Private K-means (AMOMA-DPK) clustering. Unlike many existing approaches that primarily focus on one or two objectives, the proposed method simultaneously optimizes three conflicting objectives: maximizing privacy by minimizing the total privacy budget, maximizing clustering utility by minimizing the Normalized Intra-Cluster Variance (NICV), and maximizing clustering stability by minimizing clustering instability. These objectives are formulated within a Multi-Objective Optimization (MOO) framework, enabling the generation of a diverse set of Pareto-optimal solutions that support flexible decision-making based on application-specific privacy requirements. To effectively explore the search space, AMOMA-DPK incorporates an adaptive parameter memory mechanism that dynamically adjusts crossover and mutation rates based on historically successful configurations. This is complemented by a gene-wise adaptive crossover strategy, probabilistic mutation with budget-aware perturbation, and an imbalanced local search refinement that enhances exploitation by redistributing privacy budgets across iterations. Together, these components form a robust memetic framework capable of balancing exploration and exploitation while maintaining feasibility through repair operators. Experiments on three real-world datasets show that the proposed method generally produces more competitive Pareto fronts and higher hypervolume (HV) values than representative baselines, with the highest mean HV across all tested budget and cluster-count settings on Online Shoppers. Additional comparisons with fixed allocation strategies support its effectiveness, while ablation results show that individual component effects are dataset- and budget-dependent. These findings highlight the effectiveness of AMOMA-DPK for optimizing privacy budget allocation under the evaluated settings. The source code used in this study is publicly available at https://github.com/Wayne-on-the-road/AMOMA-DPK .
Temporal knowledge graphs are usually incomplete, and representation learning techniques can be used for temporal knowledge graph reasoning. Important semantic information can be provided by entity descriptions for the learning of temporal knowledge graph representations. Most conventional approaches ignore the valuable semantic information derived from entity descriptions and only learn entity representations from structured quadruples. In this study, we offer the ED-mod temporal knowledge graph inference model, which is enhanced with entity descriptions. It can utilize the information from the graph structure and entity descriptions to perform the task of temporal knowledge graph reasoning. More specifically, by taking into account entity names and quadruple relation kinds, we first extract raw entity descriptions. Then, accurate summary data is extracted from the original entity descriptions using a unique Transformer network. Finally, we employ BiLSTM to encode entity descriptions in order to extract enough semantic information. We evaluate ED-mod on temporal knowledge graph reasoning and entity classification tasks. According to the experimental findings, our model outperforms cutting-edge baselines on the majority of criteria. The efficiency and significance of Transformer models and quadruple relation types in entity description extraction are further demonstrated by analysis of ablation experiments.
As large-scale heterogeneous graphs are difficult to label and susceptible to noisy information, Heterogeneous Graph Contrastive Learning (HGCL), as a self-supervised learning method, has been widely applied for various downstream tasks. However, traditional HGCL models decompose heterogeneous graphs into multiple homogeneous subgraphs via meta-paths without considering the influence of multi-scale properties for nodes in large graphs, which lead to incomplete feature information of nodes detected by GNN encoder. In the paper, we propose a novel Multi-scale Heterogeneous Graph Contrastive Learning (MHGCL) method with local subgraphs to capture the low-order and high-order feature information of nodes for link prediction. Specifically, we construct local subgraphs by adaptive scale module to capture specific contextual information for nodes, and then a Heterogeneous Graph Neural Network (HGNN) is employed as the encoder to learn the target node embeddings from the local view. Considering the heterogeneous graph contains indispensable semantic information, the global graph is inputted into the shared HGNN encoder to learn the deep structural information form the global view. To alleviate the label problems, we propose a dual-view contrastive loss function to train the HGNN encoder for maximizing the common information of target nodes between the local contextual information and global structural information. Finally, the link prediction experiments are employed on four real-world datasets for evaluating the performance of the proposed MHGCL and the results show that the MHGCL model is comparable to classical and state-of-the-art graph representation learning methods.
