Recent advances in real image super-resolution (Real-ISR) using diffusion models often inject low-resolution (LR) images via ControlNet, with training starting from the corresponding high-resolution (HR) images. This framework suffers from a training-inference mismatch: the noise scheduling scheme leaves residual signals even at the diffusion termination step, resulting in a non-zero signal-to-noise ratio (SNR). Consequently, the model learns to rely on residual high-frequency details from the HR input during training, details absent when inference begins from pure Gaussian noise, constraining texture fidelity. To bridge this gap, we propose a Dynamic Texture Noise Synthesizer (DTNS) that generates content-adaptive, spatially-varying texture noise from the LR image’s local gradient statistics, approximating missing high-frequency components. By embedding this synthesized noise into the initial pure noise, we provide an informed starting point for inference. A noise decoupling loss further steers synthesis away from irrelevant degradation patterns. Our method enhances perceptual quality and high-frequency reconstruction, outperforming existing techniques. This advancement is particularly beneficial for multimedia retrieval, where accurately enhancing low-quality query images can significantly improve retrieval precision and user experience.

Vision Foundation Models (VFMs) have demonstrated remarkable performance across various visual downstream tasks. To transfer their strong representation ability into lightweight CNNs, knowledge distillation serves as a promising approach. However, existing feature-based distillation methods typically perform alignment in spatial or channel dimensions, often neglecting the explicit semantic meaning of intermediate features. This semantic agnosticism leads to inefficient transfer, where students mimic the teacher’s activation patterns without grasping the underlying class-discriminative logic. In this paper, we propose Class-space Projection Distillation (CPD), a novel framework that enforces semantic alignment directly within the feature space. At the core of CPD is the Class-Projection Attention (CPA) module, which introduces a parallel dual-stream interaction mechanism: it complements standard spatial-texture retention with explicit class-semantic alignment. Specifically, CPA projects feature queries and keys into the global class space to generate Semantic Consistency Maps. Crucially, we design a versatile value projection strategy that leverages these maps for two distinct purposes: (1) Semantic-Guided Fusion, where values are projected to the feature dimension to refine student representations based on class consistency; and (2) Generative Class Activation Map (CAM) Supervision, where values are parallelly projected to the class dimension to directly generate intermediate CAMs, enforcing explicit pixel-level semantic constraints. Extensive experiments on CIFAR-100, Tiny-ImageNet, and ImageNet-1K demonstrate that CPD effectively bridges the architectural gap, consistently outperforming state-of-the-art methods.

The computational and memory demands of deep neural networks for vision tasks remain a critical barrier to their deployment on resource-constrained edge devices. Although knowledge distillation (KD) effectively transfers over-parameterized models’ knowledge into compact students, its efficacy diminishes substantially when a significant capacity gap exists between them. Current approaches often impose linear mapping constraints between output distributions, an assumption that becomes prohibitively restrictive under such capacity gaps. This paper proposes a fundamental relaxation of alignment requirements. Specifically, rather than enforcing strict parametric relationships, we experimentally validate that preserving monotonic rank correlation between teacher and student outputs suffices for effective knowledge transfer. To operationalize this insight, we introduceMonotonic Rank Knowledge Distillation, a novel framework that leverages differentiable approximations of Kendall’s rank correlation coefficient to measure and optimize rank-order consistency. Our methodology further decomposes rank correlation into inter-class and intra-class components, ensuring the student network retains both global discriminative patterns and fine-grained categorical distinctions inherent to the teacher’s outputs. Extensive experiments across CIFAR-100 and ImageNet-1K benchmarks validate the effectiveness of our approach, demonstrating consistent performance gains over state-of-the-art distillation methods. The proposed framework achieves superior generalization across diverse architectures, including CNN-based, MLP-based, and ViT-based, with particular efficacy in various compression scenarios.

Knowledge distillation has been proven to be an effective method for enhancing model performance, particularly in the domain of model compression. In this study, we propose a comprehensive approach that utilizes Brownian Distance Covariance (BDC) to measure the discrepancy between the logits produced by the teacher and student models. Unlike the conventional KL divergence used in traditional knowledge distillation, BDC captures not only linear relationships but also nonlinear dependencies, thereby overcoming the limitations of KL divergence and enabling the student model to learn more effectively from the teacher model. Additionally, our method aligns the discrepancies between the teacher and student models from both intra-class and inter-class perspectives. Extensive experiments demonstrate that our method achieves state-of-the-art (SOTA) performance across various network architectures and datasets. The code and resources related to this work are available at the following link: https://github.com/hengyin23654/BDCKD.

Post-training quantization (PTQ) has emerged as a promising approach for converting full-precision models into compact, low-precision models with minimal computational overhead, making them ideal for deployment in resource-constrained edge scenarios. While most existing PTQ techniques focus on minimizing the numerical discrepancy between model activations before and after quantization, such methods often overlook the inherent noise and distributional shifts caused by quantization, which can lead to severe performance degradation. To address this, we propose Distribution-Enhanced Reconstruction for PTQ (DEQuant), a novel approach that enhances the performance of quantized models by introducing a module that further enhances the alignment of activation pre- and post-quantization during model reconstruction. Extensive experiments demonstrate the effectiveness of DEQuant in several low-bit settings, achieving superior performance compared to existing methods. For instance, DEQuant achieves 14.18% accuracy on MobileNetV2 under the W2A2 configuration, representing a 5.72% improvement over the baseline QDrop and surpassing other baselines by 1–3%.

Deep neural networks for SAR image recognition often require compression for deployment on remote sensing platforms with limited computational and storage resources. Knowledge distillation (KD) is a key approach to improving the accuracy of lightweight networks. However, existing KD methods face challenges when applied to SAR images due to the small dataset size and the high noise in SAR images. To address this, this paper proposes a novel knowledge distillation method that relaxes the requirement for a strict linear relationship between the outputs of lightweight and large models, focusing instead on maintaining a Monotonic Relationship (MRKD). This reduces the difficulty of the KD task. Experiments on various SAR image classification and object detection datasets demonstrate that MRKD achieves state-of-the-art performance improvements for lightweight networks.

Vision Transformers (ViTs) have outperformed traditional Convolutional Neural Networks (CNN) across various computer vision tasks. However, akin to CNN, ViTs are vulnerable to backdoor attacks, where the adversary embeds the backdoor into the victim model, causing it to make wrong predictions about testing samples containing a specific trigger. Existing backdoor attacks against ViTs have the limitation of failing to strike an optimal balance between attack stealthiness and attack effectiveness.In this work, we propose an Attention Gradient-based Erosion Backdoor (AGEB) targeted at ViTs. Considering the attention mechanism of ViTs, AGEB selectively erodes pixels in areas of maximal attention gradient, embedding a covert backdoor trigger. Unlike previous backdoor attacks against ViTs, AGEB achieves an optimal balance between attack stealthiness and attack effectiveness, ensuring the trigger remains invisible to human detection while preserving the model’s accuracy on clean samples. Extensive experimental evaluations across various ViT architectures and datasets confirm the effectiveness of AGEB, achieving a remarkable Attack Success Rate (ASR) without diminishing Clean Data Accuracy (CDA). Furthermore, the stealthiness of AGEB is rigorously validated, demonstrating minimal visual discrepancies between the clean and the triggered images.

The Zero-Shot Sketch-Based Image Retrieval (ZS-SBIR) task seeks to match images with the same semantic essence as a hand-drawn sketch from a vast image repository. Given the stark contrast in information density between simple-line sketches and detailed images, this task encounters two formidable challenges: 1) Network layers focus differently on semantically relevant features across the two domains, and 2) The sparse information in sketches hampers the extraction of meaningful features. In response, we introduce the innovative Cross-Domain Feature Semantic Calibration (CD-FSC) model. This model begins by evaluating semantic correlations between domains and layers through attention map analyses in vision transformers to ensure precise semantic alignment. Subsequently, it harnesses category associations learned from the image domain to bolster semantic learning in the sketch domain. Our extensive comparative experiments across three prevalent ZS-SBIR datasets affirm that our model sets a new benchmark, outperforming current leading methods.

Recently, computation offloading methods have greatly improved the Quality of Experience (QoE) in Multi-access Edge Computing (MEC) by offloading tasks to the edge servers. Since well-coordinated actions of Terminal Devices (TDs) are critical to improving the performance of the entire individual system, many practical MEC-based applications, i.e., firefighting robots and unmanned aerial vehicles, require great teamwork among TDs. However, real-world scenarios are usually bound by resource conditions. For instance, network connectivity may weaken or experience interruptions during emergency situations. In cases where the communication medium is utilized by multiple TDs, achieving effective coordination poses a significant challenge. In this paper, we propose a computation offloading scheme based on Scheduled Multi-agent Deep Reinforcement Learning (SMDRL) to make the most efficient decision in a resource-constrained scenario. First, we design a virtual energy queue based on the MEC system and maximize the QoE (related to service delay and energy consumption) in a real-time manner. Subsequently, we propose a scheduled multi-agent deep reinforcement learning algorithm to support each TD in learning how to encode messages, select actions, and schedule itself based on the received messages. Furthermore, a TopK mechanism is introduced. This mechanism chooses the most crucial TDs to broadcast their messages, and then the computation offloading problem in a communication-constrained MEC environment can be solved in a low-communication manner. Also, we prove that even under limited communication conditions, our proposed methods can still lead to the close-to-optimal performance. The final performance analysis shows that the developed scheme has significant advantages over other representative schemes.

The data publication from multiple contributors has been long considered a fundamental task for data processing in various domains. It has been treated as one prominent prerequisite for enabling AI techniques in wireless networks. With the emergence of diversified smart devices and applications, data held by individuals becomes more pervasive and nontrivial for publication. First, the data are more private and sensitive, as they cover every aspect of daily life, from the incoming data to the fitness data. Second, the publication of such data is also bandwidth‐consuming, as they are likely to be stored on mobile devices. The local differential privacy has been considered a novel paradigm for such distributed data publication. However, existing works mostly request the encoding of contents into vector space for publication, which is still costly in network resources. Therefore, this work proposes a novel framework for highly efficient privacy‐preserving data publication. Specifically, two sampling‐based algorithms are proposed for the histogram publication, which is an important statistic for data analysis. The first algorithm applies a bit‐level sampling strategy to both reduce the overall bandwidth and balance the cost among contributors. The second algorithm allows consumers to adjust their focus on different intervals and can properly allocate the sampling ratios to optimize the overall performance. Both the analysis and the validation of real‐world data traces have demonstrated the advancement of our work.

The huge amount of graph data are published and shared for research and business purposes, which brings great benefit for our society. However, user privacy is badly undermined even though user identity can be anonymized. Graph de-anonymization to identify nodes from an anonymized graph is widely adopted to evaluate users' privacy risks. Most existing de-anonymization methods which are heavily reliant on side information (e.g., seeds, user profiles, community labels) are unrealistic due to the difficulty of collecting this side information. A few graph de-anonymization methods only using structural information, called seed-free methods, have been proposed recently, which mainly take advantage of the local and manual features of nodes while overlooking the global structural information of the graph for de-anonymization.