Yu Zhou

The aim of federated multimodal recommendation is to leverage rich multimodal information about items to predict user preferences while protecting user privacy. However, existing federated multimodal recommendation algorithms assume that the multimodal information of items held by clients is complete, whereas in real scenarios, such information is often noisy and incomplete. Traditional federated multimodal recommendation algorithms cannot be directly extended to scenarios where modalities are incomplete. To address this issue, this paper proposes a novel federated multimodal recommendation algorithm for modalities-incomplete scenarios. Specifically, to address the issue of missing modalities, we design a multi-scale residual generation network that is pre-trained on a large public dataset to reconstruct the missing modality information. Furthermore, to mitigate the imbalance in how different modalities influence the fused representation across clients, we propose an additive attention module based on noise-conditioned information entropy to dynamically guide modality fusion. Experiments on both modality-complete and modality-incomplete recommendation scenarios demonstrate the superior performance of the proposed algorithm.

In recent years, adversarial attacks in hyperspectral image (HSI) classification have attracted increasing attention. However, most existing approaches rely on the loss or gradient information of a single classifier to optimize pixel-level adversarial perturbations, which undermines the reflectance consistency of hyperspectral data and limits the transferability of adversarial examples across different classifier architectures. These drawbacks compromise both the physical interpretability and the practical applicability of adversarial examples. To address these challenges, this paper proposes an endmember extraction guided transferable adversarial attack (EEGTAA) framework for generating adversarial examples that preserve reflectance consistency while achieving strong transferability. Specifically, EEGTAA first employs endmember extraction techniques to obtain the endmember composition of HSIs and introduces adversarial perturbations into the endmember matrix, thereby ensuring the physical plausibility of the generated samples. Subsequently, EEGTAA integrates the loss gradients from multiple heterogeneous classifiers to optimize perturbations on the endmember matrix, enhancing the threat to target classifiers with unknown architectures and parameters. Furthermore, a conflict-mitigation gradient optimization strategy is theoretically derived, which utilizes weighted loss gradients to corrects the perturbation update direction within the hyperspherical neighborhood of the averaged gradient, alleviating the suboptimal optimization caused by divergent gradient directions across classifiers. Experimental results on benchmark HSI datasets demonstrate that EEGTAA not only produces adversarial examples with high attack performance and reflectance consistency but also exhibits superior transferability across classifier architectures.

Federated recommendation systems aim to provide high-quality recommendations while protecting user privacy. However, existing federated recommendation algorithms typically use a unified item embedding framework, where all clients share the same item representations. Such approaches often fail to capture users’ personalized perceptions of the same item and struggle to reflect subtle changes in user preferences, limiting the effectiveness of personalized recommendations. Moreover, statistical heterogeneity among clients poses additional challenges for model optimization. To address these issues, we propose APFMRec, a personalized federated recommendation framework that enhances both personalization and global optimization. APFMRec incorporates an adaptive item embedding module to dynamically adjust item representations based on individual user preferences. In addition, a meta-learning update module is designed to mitigate statistical heterogeneity and improve collaborative optimization across clients. Extensive experiments on five real-world datasets demonstrate the effectiveness of APFMRec. The proposed framework consistently outperforms existing federated recommendation methods, achieving up to 8.1% improvement in HR@10 and 7.8% improvement in NDCG@10.

Untargeted adversarial abundance attacks in hyper-spectral image classification have received growing attention due to their budget efficiency and physical plausibility, which arising from adherence to spectral mixing principles. However, it is difficult to design abundance-domain perturbation for targeted attacks to simultaneously suppress evidence for the source class while amplifying evidence for a specified target. This paper proposes a targeted adversarial abundance attack framework that steers reconstructed spectra toward the decision region of the designated class. The proposed method first applies deep learning based unmixing to obtain endmembers and abundances without reliance on external spectral libraries. A class-guided perturbation strategy is proposed to enforce abundance-vector similarity and feature-space alignment. Then a gradient conflict resolution mechanism is proposed to balance multiple loss objectives, which stabilizes updates and maintains small, physically consistent perturbations. Experiments on benchmark hyperspectral image datasets demonstrate that the proposed targeted adversarial abundance-attack framework achieves targeted success with improved imperceptibility and small perturbation norms.

Multitask learning (MTL) can effectively exploit the complementary information across different tasks in hyperspectral images (HSIs) to support classification. However, existing MTL methods for HSI classification usually employ shared structures in the early stages or simple information fusion in the later stages, failing to adequately characterize task-specific feature differences and adaptively select task-relevant information. Therefore, we propose a multi-stage information sharing multitask hyperspectral classification network with unmixing assistance (MSUANet). The proposed method adopts a three stage strategy, including low-level information sharing, task-specific information mining, and adaptive latent information selection, to fully exploit the information gain introduced by the unmixing task and provide more fine-grained spectral details for improved classification performance. In addition, we propose an information diverter with hybrid attention (IDHA) module, which enhances task-specific information for classification and unmixing, respectively. Furthermore, we design the adaptive latent abundance knowledge transfer (ALAKT) module, which adaptively selects latent complementary information for both tasks. Experimental results on Indian Pines, Houston 2013, Berlin, and Washington DC Mall datasets demonstrate that MSUANet outperforms state-of-the-art methods. The source code will be available at https://github.com/chihuajiao/HSIMTL.

Existing radar anti-jamming methods are difficult to effectively integrate and utilize multimodal input data, resulting in limited jamming discrimination performance in complex deceptive jamming environments. In order to achieve more sufficient information utilization, VIB-CTFN, a convolutional neural network-Transformer fusion network based on variational information bottleneck optimization is proposed, which introduces an information-theoretic approach to solve the issue of information redundancy in current fusion strategies. The network leverages complementary feature extraction. The convolutional neural network is utilized to extract the microscopic electromagnetic features of signal modality, while Transformer captures the macroscopic long-range kinematic dependencies of data modality, and the multimodal representations are fused with the feature-level multi-head attention mechanism. During alternate training, variational information bottleneck compresses redundant and irrelevant information through variational approximation, optimizes the target critical feature representation, enhancing the generalization ability and robustness of the model. Simulation experiments show that the accuracy of VIB-CTFN in complex deceptive jamming scenarios reaches 99.79%, which is 9.47% and 3.33% higher than that of convolutional neural network and Transformer without variational information bottleneck optimization, while 18.48% and 16.41% higher than that of the traditional methods signal and data fusion-based traditional interference model, demonstrating that the method has a robust and efficient jamming discrimination capability in complex electromagnetic environments.

Large language models (LLMs) have demonstrated remarkable capabilities across various natural language processing tasks, but they remain vulnerable to adversarial attacks and pose significant security concerns. Existing attack methods often treat adversarial prompts as flat sequences, neglecting the rich hierarchical structure of natural language, which could limit their effectiveness. Advancing the methodologies for adversarial attacks is crucial for rigorously assessing the security of LLMs and identifying subtle vulnerabilities. This paper introduces AdvGP, a novel framework that leverages genetic programming (GP) to generate adversarial prompts for LLMs. AdvGP exploits the inherent structural similarities between GP trees and natural language syntax to optimize the structure of harmful prompts. The framework incorporates a multi-granularity hierarchical attack strategy, specialized genetic operators that leverage an assisting LLM for depth-aware crossover and multi-level mutation, and a comprehensive fitness function integrating semantic consistency and attack effectiveness. The proposed method achieves competitive attack performance on multiple LLMs, consistently generating harmful outputs despite higher perplexity than some baselines. Ablation studies confirm the significant contributions of both LLM-aided and depth-aware mechanisms to AdvGP’s effectiveness. Furthermore, transferability analysis reveals that the generated prompts are able to bypass the defenses of various state-of-the-art LLMs, such as ChatGPT and Gemini.