Abstract Due to the continuous evolution of Android malware, machine learning-based malware detection systems face the challenge of performance degradation. To address this issue, active learning has been employed to retrain models with new labeled data. Traditionally, active learning relies on ground-truth labels, which are time-consuming to obtain. Although leveraging model-predicted pseudo-labels for model retraining offers a cost-effective alternative, incorrect pseudo-labels may lead to model self-contamination. To alleviate the annotation overhead during model retraining and mitigate the detrimental effects of erroneous pseudo-labels on active learning performance, we introduce a novel framework, PLCDroid. The framework incorporates a label correction mechanism when using pseudo-labels for model retraining. Specifically, we present a pseudo-label type recognition method (PTR) based on model uncertainty and confidence to identify incorrect pseudo-labels. On the basis of PTR, we design fine-grained correction strategies to refine pseudo-labels. Consequently, the proposed method mitigates pseudo-label errors, thereby improving malware detection performance under concept drift. Experimental results over a decade-long period demonstrate the effectiveness of our approach. In the retraining task, leveraging corrected pseudo-labels leads to a substantial performance gain. Specifically, the false negative rate decreases from 76.0% to 47.6% on average, corresponding to an improvement of 37.4% compared to the related pseudo label-based active learning method MORPH.

As the cognitive engine of urban Cyber-Physical Systems (CPS), accurate traffic prediction is essential for optimizing physical flows through cybernetic control. Cross-city traffic prediction is critical for intelligent transportation systems but faces a fundamental dilemma: purely local training suffers from data scarcity in emerging cities, while centralized training compromises data privacy. FL offers a decentralized learning alternative, yet traditionally struggles with significant communication overhead and domain heterogeneity. Existing methods typically rely on iterative, multi-round parameter synchronization to align disparate traffic patterns, which is often infeasible in bandwidth-constrained cyber–physical systems. In this paper, we propose SWIFT ( S patiotemporal W eighed I ntegration for F ederated T raffic Prediction), a communication-efficient framework that achieves high-performance traffic prediction with only One-shot of communication. Unlike methods that blindly aggregate these conflicting features, SWIFT explicitly decomposes them into orthogonal semantic subspaces. This disentanglement ensures that only semantically aligned temporal knowledge is shared, effectively eliminating the need for iterative correction. Furthermore, the framework incorporates a context-aware dynamic gating mechanism that adaptively weights and fuses these decoupled features according to real-time traffic conditions. Extensive experiments on four real-world datasets demonstrate that SWIFT matches the performance of mature multi-round FL methods, offering a scalable paradigm for data-scarce urban environments.

One-Shot Federated Learning (OSFL) enables collaborative training in a single round, eliminating the need for iterative communication, making it particularly suitable for use in resource-constrained and privacy-sensitive applications. This survey offers a thorough examination of One-Shot FL, highlighting its distinct operational framework compared to traditional federated approaches. One-Shot FL supports resource-limited devices by enabling single-round model aggregation while maintaining data locality. The survey systematically categorizes existing methodologies, emphasizing advancements in client model initialization, aggregation techniques, and strategies for managing heterogeneous data distributions. Furthermore, we analyze the limitations of current approaches, particularly in terms of scalability and generalization in non-IID settings. By analyzing cutting-edge techniques and outlining open challenges, this survey aims to provide a comprehensive reference for researchers and practitioners seeking to design and implement One-Shot FL systems, advancing the development and adoption of One-Shot FL solutions in real-world, resource-constrained settings.

Vision-Language Models (VLMs) such as CLIP have demonstrated remarkable capabilities in understanding relationships between visual and textual data through joint embedding spaces. Despite their effectiveness, these models remain vulnerable to adversarial attacks, particularly in the image modality, posing significant security concerns. Building upon our previous work on Adversarial Prompt Tuning (AdvPT), which introduced learnable text prompts to enhance adversarial robustness in VLMs without extensive parameter training, we present a significant extension by introducing the Neural Augmentor framework for Multi-modal Adversarial Prompt Tuning (NAP-Tuning). As a significant extension, NAP-Tuning first establishes a comprehensive multi-modal (text and visual) and multi-layer prompting framework. The core of this framework is a targeted structural augmentation for feature-level purification, implemented through our Neural Augmentor approach. This framework implements feature purification by incorporating TokenRefiners-lightweight neural modules that learn to reconstruct purified features via residual connections-to directly address distortions in the feature space. This structural intervention is what enables the multi-modal and multi-layer system to effectively perform modality-specific and layer-specific feature rectification. Comprehensive experiments demonstrate that NAP-Tuning significantly outperforms existing methods across various datasets and attack types. Notably, our approach shows significant improvements over the strongest baselines under the challenging AutoAttack benchmark, outperforming them by 32.3% on ViT-B16 and 31.3% on ViT-B32 architectures while maintaining competitive clean accuracy. This work highlights the efficacy of internal feature-level intervention in prompt tuning for adversarial robustness, moving beyond input-side alignment approaches to create an adaptive defense mechanism that can identify and rectify adversarial perturbations across embedding spaces.

Recent advancements in domain generalization for deepfake detection have attracted significant attention, with previous methods often incorporating additional modules to prevent overfitting to domain-specific patterns. However, such regularization can hinder the optimization of the empirical risk minimization (ERM) objective, ultimately degrading model performance. In this paper, we propose a novel learning objective that aligns generalization gradient updates with ERM gradient updates. The key innovation is the application of perturbations to model parameters, aligning the ascending points across domains, which specifically enhances the robustness of deepfake detection models to domain shifts. This approach effectively preserves domain-invariant features while managing domain-specific characteristics, without introducing additional regularization. Experimental results on multiple challenging deepfake detection datasets demonstrate that our gradient alignment strategy outperforms state-of-the-art domain generalization techniques, confirming the efficacy of our method. The code is available at https://github.com/Lynn0925/RoGA.

Android’s dominant market share has made it a prime target for malware attacks. Although machine learning-based detection systems have demonstrated effectiveness, they remain vulnerable to adversarial attacks, which modify samples to preserve malicious functionality while evading detection. Adversarial training is a prevalent defense strategy. However, generating effective adversarial examples for Android malware is challenging due to the complex mapping between feature and problem space. To address this, recent efforts have explored feature-space attacks constrained by statistical dependencies. Yet, such approaches inherently rely on large-scale datasets to achieve strong performance, and may fail to capture the underlying semantic relationships among features, like call associations. In this paper, we propose a novel method that incorporates semantic dependencies, i.e., API dependencies extracted from function call graphs of APKs. By leveraging these dependencies as domain constraints, our method preserves intrinsic call associations among features during perturbation. This leads to adversarial examples that more closely reflect realistic attack behaviors. Furthermore, a reinforcement learning-based mechanism is employed to enhance the evasive capability of the generated adversarial samples against detection models. The resulting adversarial samples are leveraged for adversarial training to enhance detector robustness. Experimental results demonstrate that the adversarial examples generated by our approach effectively enhance model robustness via adversarial training, yielding superior resilience in realistic adversarial environments. In adversarial attack scenarios, the proposed method attains the highest detection accuracy against problem-space attacks, surpassing the baseline model without adversarial training by 45.7% and 14.3%, respectively. Moreover, our method significantly reduces the average generation time by 83.5% compared to problem-space adversarial example generation approaches.