Lu Dong

While deep learning has advanced underwater image enhancement (UIE), progress is fundamentally hampered by the scarcity of paired training data. Semi-supervised methods, which leverage abundant unlabeled images, offer a promising solution but often employ naive, static weighting for pseudo-labels. This overlooks a key principle: the most informative training samples are those that represent a successful restoration from severe degradation, thus possessing a greater transformation difficulty. To address this, we propose a novel adaptive training paradigm, the Self-Distilled Degradation-Aware Network (SDDA-Net), governed by an incentive-driven weighting mechanism. We introduce a scoring function that evaluates each pseudo-label by considering both its perceptual quality and a term that is inversely proportional to its structural similarity to the input. This score, which quantifies the difficulty and success of the enhancement, is then used to dynamically modulate the sample's influence on the unsupervised loss. By allocating greater learning weights to successfully restored, complex examples, our framework forces the model to focus on mastering challenging degradations. Extensive experiments on multiple benchmark datasets demonstrate that this adaptive approach significantly outperforms current state-of-the-art methods, validating that the strategic, sample-adaptive modulation of the learning process is critical for success in data-scarce restoration tasks.

Federated learning (FL) collaboratively trains a global model across multiple clients without sharing local data, effectively utilizing data while preserving privacy. However, real-world data are often non-independently and identically distributed ( non-IID) and heterogeneous across clients, making standard FL less suitable for practical deployment. To tackle these challenges, we propose a novel framework named GBG which includes grouping, block training and global distillation. First, it introduces a novel grouping mechanism that clusters heterogeneous clients into different groups. Clients in these groups perform different training tasks. Second, it employs Symmetric Balanced Incomplete Block Design (SBIBD) to construct intra-group blocks and establishes a new training paradigm called block training. Finally, by incorporating mutual learning, the GBG framework enables the effective development of both personalized block-level models and a global model. Moreover, we demonstrate the convergence of block training in combination with existing works. Additional experiments further demonstrate that the GBG framework achieves favorable results in both testing accuracy and training error.

Sparse adversarial attacks perturb only a few pixels to achieve an attack, making them harder to detect and more dangerous. Recently, generative sparse attacks decouple the generation of sparse adversarial examples (AEs) into dense perturbations and sparse masks. By modeling the data distribution from clean examples to sparse AEs, generative sparse attacks mitigate the poor transferability that arises from over-reliance on gradients. These methods put effort into deriving optimal sparse masks on the generated perturbation. However, the quality of perturbation generation has always been overlooked, which limits the transferability of sparse AEs. To explore the influence of perturbation quality, we conduct empirical analyses of sparse gradient-based perturbations. The results show that directly applying sparsity to gradient-based perturbations disrupts their holistic adversarial information, leading to degraded attack performance. Therefore, it is critical to extract key adversarial knowledge from gradient-based perturbations while preserving their overall integrity to guide sparse adversarial attacks. Motivated by this observation, we propose to extract essential adversarial information from gradient-based AEs to guide the generator to produce higher-quality dense perturbations and stronger transferable sparse AEs. Specifically, we introduce the Gradient Perturbation Guidance (GPG) sparse adversarial attack, which integrates gradient adversarial feature guidance and gradient perturbation guidance regularization. The former guides the generator to capture gradient-based adversarial features during encoding, while the latter refines adversarial knowledge from gradient-based perturbations during decoding. Extensive experiments on ImageNet-1K show that our GPG significantly boosts transferability compared to state-of-the-art methods under consistent sparsity constraints. Our code is available at Github.

This paper investigates a dynamic beam hopping and power allocation problem in rate-splitting multiple access (RSMA) assisted multi-beam satellite system, in which beam services are allocated for terrestrial cells with non-uniform traffic demands and spectrum resource is full-frequency reused among beams for serving multiple ground terminals through RSMA technique. The objective is to jointly optimize beam hopping and power allocation strategy for maximizing the long-term sum rate while simultaneously satisfying delay fairness demand. However, the interference caused by full frequency reuse among beams limits the efficacy of RSMA, and the frequent beam patterns changes make instantaneous channel state information (CSI) detection a challenge. To address these challenges, this paper designs a novel deep reinforcement learning (DRL) based resource allocation framework for discrete-continuous action spaces. To alleviate explosion of action space, each beam is treated as a DRL agent, and a multi-agent deep reinforcement learning method is proposed to learn the beam scheduling policy. In addition, an action mapping mechanism is employed to further manage the DRL power allocation for satisfying the power input constraint arising from limited satellite power budget. Simulation results demonstrate that the proposed method can achieve the highest sum rate of the system while ensuring fairness among non-uniform traffic load cells compared with several baseline methods.