As the Segment Anything Model (SAM) becomes a popular foundation model in computer vision, its adversarial robustness has become a concern that cannot be ignored. This work investigates whether it is possible to attack SAM with image-agnostic Universal Adversarial Perturbation (UAP). In other words, we seek a single perturbation that can fool the SAM to predict invalid masks for most (if not all) images. We demonstrate conventional image-centric attack framework is effective for image-dependent attacks but fails for universal adversarial attacks. To this end, we propose a perturbation-centric framework for UAP generation on SAM, which shifts the optimization focus from image-specific clean and adversarial image pairs to the UAP itself. Based on this framework, we develop a contrastive learning method in which the UAP is treated as the anchor sample to be optimized, its augmented version is used as the positive sample, and natural images are used as negatives. The effectiveness of our proposed perturbation-centric UAP generation method is validated by both quantitative and qualitative results. On top of the ablation study to understand various components in our proposed method, we shed light on the roles of positive and negative samples in making the generated UAP effective for attacking SAM.
Predicting future cellular network traffic volume patterns is crucial for optimizing network resource management and enhancing user experience. Recently, with cellular traffic data represented as pseudo-image data, state-of-the-art frameworks based on Deep Neural Networks (DNNs) have been introduced to enable effective modeling of spatiotemporal dependencies for the prediction tasks. Nonetheless, highly parameterized DNNs require large datasets, which are often missing for cellular network traffic prediction task and thus underscore the need for effective data augmentation strategies. Our preliminary analysis shows that the augmentation methods for natural images prove ineffective in enhancing performance on this task due to its pseudo-image nature. We identify that the reason lies in image augmentations introducing missing values and misalignments, with masking and geometric alterations disrupting the real-world cellular traffic patterns. On top of this, we propose an adaptation of underlying augmentation process into a tailored strategy suitable for the spatial and temporal complexities of the data, which stem from interactions between base stations and evolving traffic patterns. Specifically, we introduce MixScale, an augmentation technique designed for pseudo-image cellular network traffic data. MixScale integrates constrained spatio-temporal data mixing with multi-scaling to better align with the unique characteristics of the dataset. Evaluation on two real-world datasets, including a 5G dataset, demonstrates promising results; MixScale consistently achieves reductions between 16% and 32% in root mean square error (RMSE) compared to the baseline.
This study incorporates the concept of infrastructure (evacuation resource) into the existing PRISM simulation framework to optimize infrastructures allocation during radiological emergencies. A reinforcement learning methodology is applied to derive spatiotemporal allocation strategies for these infrastructures by explicitly reflecting the dynamic movements of evacuees. The reinforcement learning-optimized infrastructure distributions reduce evacuation completion time compared to uniform and population-proportional allocations, thereby providing decision makers with an efficient infrastructure allocation strategy that accounts for both dynamic evacuation conditions and limited infrastructure availability.
Sustainable mass production of high-quality graphene is essential for its use as a conductive additive in high-power energy storage systems, as conventional synthesis often causes structural damage and environmental toxicity. Here, we report an eco-friendly anodic electrochemical exfoliation strategy using a dual-salt aqueous electrolyte of NaCl and NaClO4. This synergistic approach exploits the small hydrated radius of chloride ions (Cl−) for initial channel opening and the strong expansion effect of perchlorate ions (ClO4−) for rapid delamination, preserving the carbon lattice integrity. By optimizing the NaCl/NaClO4 molar ratio, we balanced defect density and carbon purity to maximize battery performance. The optimized process yielded high-quality few-layer graphene with carbon content exceeding 94 at%, low defect density (ID/IG = 0.13), and high powder electrical conductivity (1237.9 S/cm). Used as a two-dimensional (2D) conductive additive in thick, high-loading graphite anodes (∼4.5 mg/cm2) at a reduced content of 0.55 wt%, the graphene enabled 82.7% capacity retention at an ultrahigh rate of 15C and 99.6% retention after 100 cycles at 0.5C via a robust 2D conductive network. These results confirm that the crystalline 2D network facilitates Li-ion transport and mechanical stability, highlighting the potential of this green route for next-generation high-power lithium-ion batteries.
