Cyber-physical-social systems (CPSS) generate heterogeneous data in which decision-critical events, such as cyber intrusions, equipment faults, social anomalies, and minority disease patterns, are often rare. This class imbalance undermines the reliable recognition required by cyber-physical-social intelligence (CPSI). Existing oversampling pipelines commonly rely on a fixed 1:1 balance ratio, emphasize minority-distribution fidelity while paying limited attention to inter-class separability, and usually produce deterministic decisions without modeling the uncertainty introduced by synthetic samples. To address these issues, we propose an Inter-class Maximizing GAN with Multiratio Oversampling (IMGAN-MO). First, an inter-class distribution maximization mechanism is embedded into the generative adversarial networks (GAN) objective, where a Wasserstein-distance-based separation term encourages the augmented minority distribution to remain faithful to minority semantics while increasing its margin from the majority class. Second, a multiratio oversampling strategy constructs augmented datasets across a range of imbalance ratios and trains an ensemble of classifiers. Finally, a Dempster-Shafer (DS) evidence fusion scheme dynamically discounts each classifier according to its local consistency around the test sample, yielding an uncertainty-aware decision. Experiments on twelve imbalanced benchmark datasets show that IMGAN-MO achieves the best or competitive performance in most cases compared with classical oversampling and deep generative baselines, indicating its potential as a robust component for data-driven CPSI services.
Cascade popularity prediction is vital in information diffusion studies. Existing models either use stochastic processes or deep neural networks to simulate or predict cascade growth patterns, but they often overlook the role of inter-cascades in propagation, like competition, cooperation, and star user authority. Consequently, these models can’t capture intra-cascade growth and inter-cascade impact at the same time, causing poor prediction performance. To effectively track both internal and external properties of the cascade growth process, models must consider both growth characteristics and inter-cascade effects. To address the above problem, we propose a new framework in the end-to-end manners, which can simultaneously model the intra and inter properties of information cascades. In particular, we first propose a cascade self-enhancement method to simply and effectively extract the intrinsic characteristics of the intra-cascade growth process. Second, rather than focusing on single cascade, we capture the effect of interactions among cascades on cascade diffusion from a global perspective via analyzing the diffusion characteristics of inter-cascades. Finally, we fuze the two parts in an end-to-end model based on contrastive learning to predict the cascade popularity. We conduct extensive experiments on three large real-world datasets and the results demonstrate the superiority to state-of-the-art methods.
Rumor intervention in online social networks requires both an expressive diffusion model and a scalable decision mechanism for selecting a small set of users for truth injection. Existing diffusion-based methods often describe users with coarse belief states and assume fixed spreading willingness, which limits their ability to capture attention decay, stimulus-driven reactivation, and belief correction during rumor-truth competition. To address this problem, this article proposes seven-state truth–rumor competition (STRC), a behavior-aware competitive diffusion framework for rumor containment (RC). STRC separates users’ belief orientation from their dissemination intention and models threshold-based activation, interest-driven state transitions, conservative spreading, termination, and truth-induced correction in a unified propagation process. Based on the resulting dynamics, we formulate the RC problem and analyze its computational hardness. We further develop IScore, an interpretable cascade-based seed-selection algorithm that derives activation-time rumor-support scores from a realized rumor-only STRC diffusion and penalizes redundant truth-seed deployment. Experiments on synthetic networks, real-world social networks, and Twitter15/Twitter16 rumor cascades show that STRC can represent competitive diffusion behavior and that IScore achieves competitive containment performance with substantially lower computational cost than simulation-based greedy methods. These results indicate that the proposed framework provides a scalable and interpretable solution for behavior-aware rumor intervention in online social networks.
Federated learning (FL) breaks down data silos through distributed collaboration. Due to the limited motivation of participants, incentive mechanisms are essential for encouraging sustained participation. However, existing incentive mechanisms are vulnerable to free-rider attacks, particularly in differentially private FL (DPFL), where the injected privacy noise obscures intrinsic gradient characteristics and significantly degrades the effectiveness of existing free-rider detection methods. Moreover, most existing incentive mechanisms adopt conventional aggregation strategies without incorporating participant contributions, thereby limiting the performance of the global model. To address these challenges, this article proposes an efficient incentive mechanism against free-rider attacks for DPFL, called incentive mechanism against free-rider (IMFR). Specifically, the central server detects free-riders by performing linear regression on the cosine similarities of consecutive local updates and records their historical behaviors. Based on the historical free-riding behavior, local data size, and privacy budget of each participant, IMFR evaluates participant contributions and performs contribution-weighted model aggregation to improve the quality of the global model. Theoretical analysis demonstrates that IMFR satisfies long-term incentive compatibility, individual rationality, budget feasibility, computational efficiency, low storage overhead, differential privacy guarantee, and aggregation stability. Extensive experiments on two public datasets demonstrate that IMFR consistently achieves superior free-rider detection accuracy, incentive effectiveness, and aggregation performance compared with existing methods.
Heterogeneous graphs, with their rich structural and semantic information, provide enhanced representational power for fraud detection (FD). However, existing methods often rely on manually designed meta-paths or uniform aggregation strategies, which fail to effectively integrate information from multiple node types. This can lead to embedding distortion and reduced detection performance. In addition, heterogeneous graphs in FD, due to their complex structures and dense information, are prone to privacy leakage and face severe privacy risks. To address these challenges, this article proposes a differentiated privacy separated heterogeneous graph fraud detection (DP-SHF) framework. Under our proposed framework, we first design a type-aware feature expansion and separated aggregation mechanism to effectively integrate semantics from different node types, reducing embedding distortion and improving detection performance. Second, we develop type-aware differentiated privacy strategies by considering heterogeneous characteristics for edge-level privacy and node-level privacy; the node-level budget allocation is driven by type-level aggregate statistics that we treat as public side information, while all per-node data remain protected. We conduct experiments on four benchmarks: three multirelational fraud datasets brought into the framework by virtual type construction, and TwiBot-22, which is heterogeneous by data origin. The results show that DP-SHF achieves the best average performance across the four in FD. Additionally, with respect to privacy protection, our model still outperforms all baselines on the three multirelational benchmarks. Its average area under the receiver operating characteristic curve (AUC) degradation is much smaller than that of the edge-level and node-level baselines, and is reduced by about $\mathbf{7.7\%}$ and $\mathbf{7.6\%}$ relative to the strongest prior method in each setting.
