Contrastive learning has greatly enhanced the performance of self-supervised visual representation learning, especially in linear probing evaluations, where its accuracy now closely matches that of supervised learning on the ImageNet dataset. However, with the increasing use of the Vision Transformer, masked image modeling has emerged as an alternative self-supervised learning technique. It has shown performance that not only matches but often outperforms contrastive learning across a variety of visual tasks. This paper highlights several limitations in current contrastive learning solutions that utilize masked images, including suboptimal masking strategies, the absence of dynamic adjustments in the number of masked patches, and inefficient data augmentation techniques. To address these challenges, we propose a novel approach that introduces guided masking strategies, dynamically determines the number of masked patches, and explores lighter data augmentation during pre-training. Compared with state-of-the-art methods, our method achieves a 2.74% increase in linear probing accuracy on CIFAR-10, 2.15% on CIFAR-100, 3.47% on Tiny-ImageNet, and 0.33% on ImageNet100. Code is available at: https://github.com/Leon-jcH/RFMaCo.

Precise image manipulation localization requires a synergistic integration of global semantic context and fine-grained structural inconsistencies. However, existing refinement paradigms often struggle to effectively reconcile these multi-scale clues, resulting in suboptimal boundary adherence and false activations in authentic regions. In this paper, we propose the Structure-Aware Adaptive Refinement Network (SAR-Net) for precise image manipulation detection and localization. Our method explores manipulation-related structural semantics as a structure-aware prior to guide the decoding process. Specifically, we first introduce an Edge Prior Guidance (EPG) module to establish a global structural prior, providing reliable spatial guidance for mask reconstruction. Building upon this prior, we incorporate a Confidence Gating strategy within our cascaded decoder. Instead of passively propagating coarse predictions, this module evaluates the inherited masks against high-resolution forensic evidence, enabling the network to adaptively rectify spatial errors and suppress unreliable activations. Furthermore, a Unified Attention Module (UAM) is integrated to capture discriminative traces across both spectral and spatial domains. Extensive experiments on five benchmark datasets demonstrate that SAR-Net achieves superior localization accuracy and boundary precision, significantly outperforming state-of-the-art methods in terms of both robustness and effectiveness.

Copyright protection in Federated Learning (FL) environments faces some challenges, such as fidelity, validity, robustness, and persistence, due to the large number of diverse FL clients. To address these issues, we propose a Grouped Clients Hierarchical Watermarking (GCHW) scheme, which consists of client grouping, watermark generation, watermark embedding and watermark verification. The client grouping is to group clients based on shared characteristics. In watermark generation, a hierarchical framework is proposed to establish dependencies between the group and client watermarks by embedding part of the group watermark information into each client water mark. Moreover, our GCHW scheme innovatively embeds binary watermarks into the frequency domain of model parameters at both the group and client levels during the watermark embedding phase. Experimental results demonstrate that GCHW effectively maintains fidelity and the validity of the watermarks is validated across independent and identically distributed (IID) and non-IID scenarios. GCHW also exhibits strong robustness against common attacks and defensive aggregation algorithms. In addition, GCHW shows good persistence in dynamic FL environments.

Speech has been widely used for Parkinson’s Disease (PD) detection. However, cross-lingual PD detection is still a challenge, in which detection performance is often affected by lingual variations in cross-lingual scenarios. In other words, the detection model trained in source language domain is not suitable in target language domain. In order to alleviate this issue, a Cross-lingual PD detection method based on double-level optimized Meta-Learning (CPDML) is proposed in this paper, which includes double-level optimized Meta-training modul, adversarial data splitting module and Meta-testing module. In Meta-training phase, a domain adaptation training strategy based on double-level optimized Meta-learning is designed, which focuses on the intra-domain alignment. Adversarial data splitting module is designed to solve the inter-domain gap that exists between the support set and the query set on source domain dataset. It adversarially identifies the support set and the query set with the largest distribution difference and trains the model to generalize well under this splitting. In Meta-testing phase, a small amount of samples in target language domain are used to fine-tune the PD detection model trained in source language domain. Experimental results on the public and our self-collected PD speech datasets show that the proposed method is superior to existing methods in cross-lingual scenarios.

Drug-Target Interaction (DTI) prediction is a foundational task in computational drug discovery. While emerging deep learning-based approaches typically employ dual-tower architectures to distill intrinsic biochemical attributes, they overlook the fact that the identity of a drug or target is defined not only by its molecular structure but also by its global topological role within the pharmacological interaction space. Existing methods to incorporate relational context either remain restricted to local pairwise dependencies or rely on similarity-based graph construction, which is frequently noisy and biologically unreliable. To address these limitations, we propose TopoDTI, a topological role-aware hybrid framework that synergistically integrates global topological dependencies with intrinsic biochemical attributes without relying on auxiliary knowledge bases. Specifically, TopoDTI comprises two core components: a Global Role Encoder (GRE) and a Harmonized Hybrid Optimization (HHO) strategy. The GRE manifests latent topological roles by modeling high-order relational patterns within the DTI bipartite graph. Through a semi-supervised contrastive learning paradigm, it captures informative structural knowledge beyond immediate connectivity, yielding expressive topology-aware representations. To mediate the optimization imbalance inherent in learning from these disparate feature spaces, we introduce the HHO strategy. By decoupling early-stage optimization and progressively coordinating the distinct learning dynamics of topological and biochemical views, HHO ensures stable convergence and prevents representation collapse. Extensive experiments on benchmark datasets demonstrate that TopoDTI consistently outperforms state-of-the-art in Accuracy and AUPR. Further analyses confirm that modeling global interaction topology substantially enhances model generalization and robustness, particularly in data-sparse scenarios.

Deep neural networks (DNNs) have been shown to be vulnerable to adversarial examples crafted by adversaries to deceive the target model. Two popular approaches to mitigate this issue are adversarial training and adversarial example detection. Adversarial training aims to enable the target model to accurately recognize adversarial examples in image classification tasks; however, it often lacks generalizability. Conversely, adversarial detection demonstrates good generalization but does not assist the target model in recognizing adversarial examples. In this paper, we first define the label recovery task to address the adversarial challenges faced by DNNs. We then propose a novel generative classifier specifically for the adversarial example label recovery task. This method is termed Detection and Label Recovery (DLR), which comprises two components: Detector and Recover. The Detector processes both legitimate and adversarial examples, while the Recover component seeks to ascertain the ground-truth label of the detected adversarial example. DLR effectively combines the strengths of adversarial training and adversarial example detection. Experimental results demonstrate that our method outperforms several state-of-the-art approaches..

The neural network model has the drawback of making incorrect predictions under the influence of slight adversarial perturbations. Single-step adversarial training (AT) is an effective tool to bring adversarial robustness for the model to resist such attack. From the perspective of perturbation setting in training, we identify a conflict between the pursuit of greater robustness and the need to prevent catastrophic overfitting within the AT framework. To get out of this dilemma, we delve into the impact of perturbations on human visual perception. Our analysis reveals that examples containing more misleading features should be assigned a smaller perturbation magnitude to preserve subtle yet significant features. Conversely, examples encompassing more relevant features should be assigned a larger perturbation magnitude, enabling the model to adapt to stronger attacks effectively. Motivated by these insights, we propose a concise refinement to the AT framework to unleash its full potential for single-step AT. Instead of employing a fixed perturbation magnitude, we introduce a “band” of magnitudes, allowing each example to select an appropriate magnitude based on its visual characteristic. Through extensive experiments conducted on three datasets, we demonstrate the efficacy of our proposed strategy. Our approach not only improves the model’s robustness and prevents catastrophic overfitting but also effectively mitigates robust overfitting—an issue that has remained unresolved in the context of single-step AT, marking a significant advancement in the field.

Although deep neural networks (DNNs) have made groundbreaking progress in various machine learning domains, their practical implementation is still significantly impeded by adversarial vulnerability [ 1]. Adversarial training, the primary approach to enhance the adversarial robustness of DNNs, augments the training set with adversarial examples and applies adversarial regularization loss to improve robustness [ 2]. However, finding models that achieve a reasonable trade-off between accuracy and robustness remains an unresolved challenge. In this paper, we propose the adoption of global probability constraints to stabilize model decision-making. Our contributions can be summarized as follows.

Named entity recognition is a fundamental task in natural language processing. Recognizing entities in biomedical text, known as the BioNER, is particularly crucial for cutting-edge applications. However, BioNER poses greater challenges compared to traditional NER due to (1) nested structures and (2) category correlations inherent in biomedical entities. Recently, various BioNER models have been developed based on region classification or large language models. Despite being successful, these models still struggle to balance handling nested structures and capturing category knowledge. We present a novel parallel BioNER model, Bean, designed to address the unique properties of biomedical entities while achieving a reasonable balance between handling nested structures and incorporating category correlations. Extensive experiments on five public NER datasets, including four biomedical datasets, demonstrate that Bean achieves state-of-the-art performance. The proposed Bean is elaborately designed to achieve two key objectives of the BioNER task: clearly detecting entity boundaries and correctly classifying entity categories. It is the first BioNER model to handle nested structures and category correlations in parallel. We exploit head, tail, and contextualized features to efficiently detect entity boundaries via a triaffine model. To the best of our knowledge, we are the first to introduce a multi-label classification model for the BioNER task to extract entity category information without boundary guidance.

To comprehensively capture the diversity of articulatory impairments, multi-type speech data, such as sustained vowels, repeated syllables and text reading, should be used together for Parkinson's Disease (PD) detection. However, in real applications, it is common that one or more types of speech data are unavailable or incomplete, which severely reduces the robustness and generalization ability of the PD detection model. To alleviate this issue, in this paper, a PD detection framework based on knowledge distillation is proposed. The proposed framework consists of a teacher network and a student network. The teacher network is trained on speech data with complete types, and provides multi-level supervision information to the student network. The types of speech data for student network are incomplete or missing. The teacher and student network all include three modules, i.e., feature extraction module, gated mechanism module, and multi-scale representation module. Especially, the gated mechanism module adopts an attentionbased gated neural fusion mechanism to adaptively adjust the contribution weights of each type of speech data. Moreover, hierarchical mutual information maximization mechanism, central moment discrepancy loss and output-level distillation constraints are employed to ensure robust PD detection performance even in the presence of missing data types. The experimental results on Chinese and Italian PD speech datasets show that the proposed framework exhibits strong robustness and high accuracy under different data types missing scenarios.

The real-world datasets used for training machine learning models may contain sensitive features such as gender, race, and age, raising concerns about unfair treatment of specific groups or individuals. These concerns can frequently be alleviated by feature selection techniques. However, most existing research primarily focuses on removing statistically correlated features with sensitive attributes, which frequently fails to address the underlying causal relationships that contribute to unfair treatment. In this paper, we propose a novel fairness-aware causal feature selection method (FCFS). By constructing a causal diagram and leveraging Markov blanket theory, we identify the discrimination-prone feature set that undermine the fairness of model decisions, then propose direct and indirect causal discriminatory effect measures for each feature within the set. Features with high discrimination effect are iteratively assigned priority for removal. Experimental results on public datasets, including Adult, Boston, and German, demonstrate that the proposed method, FCFS, achieves an ideal trade-off between accuracy and fairness (measured by EOs, DP, EO, and BAD) compared to existing fairness-aware feature selection methods.

The advancement of Large Vision-Language Models (LVLMs) has propelled their application in the medical field. However, Medical LVLMs (Med-LVLMs) encounter factuality challenges due to modality misalignment, where the models prioritize textual knowledge over visual input, leading to hallucinations that contradict information in medical images. Previous attempts to enhance modality alignment in Med-LVLMs through preference optimization have inadequately addressed clinical relevance in preference data, making these samples easily distinguishable and reducing alignment effectiveness. In response, we propose MMedPO, a novel multimodal medical preference optimization approach that considers the clinical relevance of preference samples to enhance Med-LVLM alignment. MMedPO curates multimodal preference data by introducing two types of dispreference: (1) plausible hallucinations injected through target Med-LVLMs or GPT-4o to produce medically inaccurate responses, and (2) lesion region neglect achieved through local lesion-noising, disrupting visual understanding of critical areas. We then calculate clinical relevance for each sample based on scores from multiple Med-LLMs and visual tools, enabling effective alignment. Our experiments demonstrate that MMedPO significantly enhances factual accuracy in Med-LVLMs, achieving substantial improvements over existing preference optimization methods by 14.2% and 51.7% on the Med-VQA and report generation tasks, respectively. Our code are available in https://github.com/aiming-lab/MMedPO}{https://github.com/aiming-lab/MMedPO.

In recent years, despite the remarkable success of deep neural networks across diverse tasks, their vulnerability to adversarial attacks has raised significant concerns. These attacks introduce imperceptible perturbations to input samples, misleading model predictions and revealing a lack of robustness. This challenge is particularly pronounced in black-box scenarios, which more closely reflect real-world deployment settings. Among these, transfer-based attacks have garnered considerable attention due to their ability to operate without access to the target model's architecture or parameters. However, their effectiveness remains limited, primarily due to overfitting on surrogate models. To address this limitation, we propose an efficient black-box transfer-based attack method, termed Frequency-Diversified and Model-Perturbed Attack (FDMPA), which integrates frequencydomain data augmentation with model weight perturbation. FDMPA jointly optimizes both the input and model perspectives to enhance the transferability of adversarial examples. On the input side, we leverage the Discrete Cosine Transform to project samples into the frequency domain, where we apply mid-frequency interpolation mixing and low- and high-frequency masking to enrich diversity. On the model side, inspired by model pruning techniques, we inject random perturbations into the surrogate model's convolutional kernels, approximating a multi-model ensemble to alleviate overfitting and improve the generalization of adversarial perturbations. Experimental results demonstrate that FDMPA significantly improves the success rate of black-box transfer-based attacks without incurring substantial computational overhead, highlighting its practical value and scalability. Code is available at https://github.com/WangZhirui2001/FDMPA.

Deep neural networks have demonstrated superior performance in various computer vision tasks.How-ever,they have been found to be highly susceptible to adversarial attacks,which involve the addition of perturba-tions to examples during the inference phase that are imperceptible to the human eye.To defend against adversarial attacks,some works have explored the reverse engineering of adversarial examples,known as the adversarial attri-bution problem.By attributing the attack algorithm and victim model used to generate adversarial examples,de-fenders can gain insights into the attacker's knowledge and targets,thereby enabling the design of more effective defense algorithms against corresponding attacks.Existing methods have mostly approached the adversarial attribu-tion problem as a single-task learning problem.However,as the scope of attack algorithms and victim models has expanded,single-task learning has faced the challenge of combinatorial explosion.To improve the accuracy of ad-versarial attribution and meet the requirements for different attribution granularities,attack algorithms and victim models were layered,and the dependencies between different levels were utilized.A multi-task adversarial attribu-tion method based on a hierarchical structure was proposed.This method simultaneously performed the attribution tasks of attack algorithms and victim models at different levels and employed hierarchical path prediction to learn the dependencies between these levels.Experimental results on multiple datasets demonstrate that the proposed method achieves better attribution performance compared to other attribution methods.

Currently, deep learning models have gained significant popularity in time series forecasting within industrial systems due to their high accuracy. However, these models exhibit vulnerability to adversarial attacks, posing significant cost and security risks. Existing attack methods for time series, primarily adapted from those developed for image classifiers, fail to effectively explore the vulnerability of time series forecasting models, since they overlook the distinct characteristics and temporal patterns inherent in time series data. To address this challenge and inspire future research aimed at improving the reliability of time series forecasting models, we identify the goals of adversarial attacks for time series forecasting and propose a novel white-box adversarial attack method named TCA. Specifically, TCA exploits gradient information from the target model, iteratively applies perturbations to the original samples, and constrains these perturbations based on the temporal characteristics. Extensive experiments on multiple DL models and real-world time series datasets reveal the shortcomings of existing attacks for time series forecasting and demonstrate the effectiveness, stealthiness, and rationality of TCA attacks in both untargeted and targeted attack scenarios.

The increasing complexity of software systems has made them more prone to bugs, prompting the development of automated bug localization techniques to ensure software reliability. Despite these techniques having demonstrated notable success at the file level, their application and optimization at the function level often encounter serious performance cliffs. This limitation underscores the urgent need for a dedicated framework for function-level bug localization, which we address through FUNLoc, a novel framework that takes coarse-grained source files as input units and identifies fine-grained buggy functions as output. To address the critical challenges of handling domain-specific bug reports and managing vast function-level sample space, we introduce two key innovations that are seamlessly integrated into FUNLoc. First, we design a contrastive learning-based domain-adaptive language model to enhance the framework's ability to process and interpret specialized bug reports effectively. Second, we propose an active learning-based dynamic negative sampling strategy to address the scalability issues arising from the extensive function-level sample space. To evaluate the effectiveness of our approach, we extend and release a function-level bug localization dataset derived from large-scale real-world projects. Extensive experiments demonstrate that our approach outperforms state-of-the-art techniques.

Deep neural networks have advanced significantly in the last several years and are now widely employed in numerous significant real-world applications. However, recent research has shown that deep neural networks are vulnerable to backdoor attacks. Under such attacks, attackers release backdoor models that achieve satisfactory performance on benign samples while behaving abnormally on inputs with predefined triggers. Successful backdoor attacks can have serious consequences, such as attackers using backdoor generation methods to bypass critical face recognition authentication systems. In this paper, we propose PBADT, a precise backdoor attack with dynamic trigger. Unlike existing work that uses static or random trigger masks, we design an interpretable trigger mask generation framework that places triggers at positions that have the most significant impact on the prediction results. Meanwhile, backdoor attacks are made more efficient by using forgettable events to improve the efficiency of backdoor attacks. The proposed backdoor method is extensively evaluated on three face recognition datasets, LFW, CelebA, and VGGFace, while further evaluated on two general image datasets, CIFAR-10 and GTSRB. Our approach achieves almost perfect attack performance on backdoor data.

As artificial intelligence (AI) systems are deployed at scale in real world, there is a growing concern about the harmful biases when addressing target tasks. Such biases originate from unbalanced training data with respect to sensitive attributes, potentially resulting in unfair treatment of specific groups or individuals. To address these challenges, we design a fairness-aware causal diffusion model (FCDM) that leverages neural network-based architectures to generate counterfactual samples. These counterfactual samples are used to augment original dataset to achieve a balance on sensitive attributes for fair model training. Generally, a qualified counterfactual sample generation method in fair model training requires the ability to flip specific sensitive attributes of original sample while keeping other attributes unchanged as much as possible, and addresses the scenarios involving one or multiple sensitive attributes. To achieve this, FCDM identifies the set of attributes that need to be flipped simultaneously, and calculate corresponding attributes weights by causal reasoning. The chosen attributes and their weight will be input into conditional guided diffusion model to generate desired counterfactual samples by adjusting model gradient. The experiments demonstrate that the counterfactual samples generated by FCDM exhibit better validity, similarity, and diversity compared to existing methods. Moreover, the models trained on the augmented dataset generated by FCDM show ideal fairness and accuracy.

In contrastive self-supervised learning, positive samples are typically drawn from the same image but in different augmented views, resulting in a relatively limited source of positive samples. An effective way to alleviate this problem is to incorporate the relationship between samples, which involves including the top-K nearest neighbors of positive samples. However, the problem of false neighbors (i.e., neighbors that do not belong to the same category as the positive sample) is an objective but often overlooked challenge due to the query of neighbor samples without supervision information. In this paper, we present a simple self-supervised learning framework called Mixed Nearest-Neighbors for Self-Supervised Learning (MNN). MNN optimizes the influence of neighbor samples on the semantics of positive samples through an intuitive weighting approach and image mixture operations. The results demonstrate that MNN exhibits exceptional generalization performance and training efficiency on four benchmark datasets.

Contrastive learning has emerged as an essential approach in self-supervised visual representation learning. Its main goal is to maximize the similarities between augmented versions of the same image (positive pairs), while minimizing the similarities between different images (negative pairs). Recent studies have demonstrated that harder negative samples, i.e., those that are more challenging to differentiate from the anchor sample perform a more crucial function in contrastive learning. However, many existing contrastive learning methods ignore the role of hard negative samples. In order to provide harder negative samples for the network model more efficiently. This paper proposes a novel feature-level sample sampling method, namely sampling synthetic hard negative samples for contrastive learning (SSCL). Specifically, we generate more and harder negative samples by mixing them through linear combination and ensure their reliability by debiasing. Finally, we execute weighted sampling of these negative samples. Compared to state-of-the-art methods, our method can provide more high-quality negative samples. Experiments show that SSCL improves the classification performance on different image datasets and can be readily integrated into existing methods.