Efficient long-context inference is critical as large language models (LLMs) adopt context windows of ranging from 128K to 1M tokens. However, the growing key-value (KV) cache and the high computational complexity of attention create significant bottlenecks in memory usage and latency. In this paper, we find that attention in diverse long-context tasks exhibits sparsity, and LLMs implicitly "know" which tokens can be dropped or evicted at the head level after the pre-filling stage. Based on this insight, we propose Self-Attention Guided Eviction~(SAGE-KV), a simple and effective KV eviction cache method for long-context inference. After prefilling, our method performs a one-time top-k selection at both the token and head levels to compress the KV cache, enabling efficient inference with the reduced cache. Evaluations on LongBench and three long-context LLMs (Llama3.1-8B-Instruct-128k, Llama3-8B-Prolong-512k-Instruct, and Qwen2.5-7B-Instruct-128k) show that SAGE-KV maintains accuracy comparable to full attention while significantly improving efficiency. Specifically, SAGE-KV achieves 4x higher memory efficiency with improved accuracy over the static KV cache selection method StreamLLM, and 2x higher memory efficiency with better accuracy than the dynamic KV cache selection method Quest.

Embedding models have become essential tools in both natural language processing and computer vision, enabling efficient semantic search, recommendation, clustering, and more. However, the high memory and computational demands of full-precision embeddings pose challenges for deployment in resource-constrained environments, such as real-time recommendation systems. In this work, we propose a novel finetuning framework to ternary-weight embedding models, which reduces memory and computational overhead while maintaining high performance. To apply ternarization to pre-trained embedding models, we introduce self-taught knowledge distillation to finalize the ternary-weights of the linear layers. With extensive experiments on public text and vision datasets, we demonstrated that without sacrificing effectiveness, the ternarized model consumes low memory usage and has low latency in the inference stage with great efficiency. In practical implementations, embedding models are typically integrated with Approximate Nearest Neighbor (ANN) search. Our experiments combining ternary embedding with ANN search yielded impressive improvement in both accuracy and computational efficiency. The repository is available at here.

Can we directly visualize what we imagine in our brain together with what we describe? The inherent nature of human perception reveals that, when we think, our body can combine language description and build a vivid picture in our brain. Intuitively, generative models should also hold such versatility. In this paper, we introduce BrainDreamer, a novel end-to-end language-guided generative framework that can mimic human reasoning and generate high-quality images from electroencephalogram (EEG) brain signals. Our method is superior in its capacity to eliminate the noise introduced by non-invasive EEG data acquisition and meanwhile achieve a more precise mapping between the EEG and image modality, thus leading to significantly better-generated images. Specifically, BrainDreamer consists of two key learning stages: 1) modality alignment and 2) image generation. In the alignment stage, we propose a novel mask-based triple contrastive learning strategy to effectively align EEG, text, and image embeddings to learn a unified representation. In the generation stage, we inject the EEG embeddings into the pre-trained Stable Diffusion model by designing a learnable EEG adapter to generate high-quality reasoning-coherent images. Moreover, BrainDreamer can accept textual descriptions (e.g., color, position, etc.) to achieve controllable image generation. Extensive experiments show that our method significantly outperforms prior arts in terms of generating quality and quantitative performance.

Autonomous vehicles and robots often struggle with reliable visual perception at night due to the low illumination and motion blur caused by the long exposure time of RGB cameras. Existing methods address this challenge by sequentially connecting the off-the-shelf pretrained low-light enhancement and deblurring models. Unfortunately, these methods often lead to noticeable artifacts (\eg, color distortions) in the over-exposed regions or make it hardly possible to learn the motion cues of the dark regions. In this paper, we interestingly find vision-language models, \eg, Contrastive Language-Image Pretraining (CLIP), can comprehensively perceive diverse degradation levels at night. In light of this, we propose a novel transformer-based joint learning framework, named DAP-LED, which can jointly achieve low-light enhancement and deblurring, benefiting downstream tasks, such as depth estimation, segmentation, and detection in the dark. The key insight is to leverage CLIP to adaptively learn the degradation levels from images at night. This subtly enables learning rich semantic information and visual representation for optimization of the joint tasks. To achieve this, we first introduce a CLIP-guided cross-fusion module to obtain multi-scale patch-wise degradation heatmaps from the image embeddings. Then, the heatmaps are fused via the designed CLIP-enhanced transformer blocks to retain useful degradation information for effective model optimization. Experimental results show that, compared to existing methods, our DAP-LED achieves state-of-the-art performance in the dark. Meanwhile, the enhanced results are demonstrated to be effective for three downstream tasks. For demo and more results, please check the project page: \url{https://vlislab22.github.io/dap-led/}.

Backdoor attacks present significant threats to Large Language Models (LLMs), particularly with the rise of third-party services that offer API integration and prompt engineering. Untrustworthy third parties can plant backdoors into LLMs and pose risks to users by embedding malicious instructions into user queries. The backdoor-compromised LLM will generate malicious output when and input is embedded with a specific trigger predetermined by an attacker. Traditional defense strategies, which primarily involve model parameter fine-tuning and gradient calculation, are inadequate for LLMs due to their extensive computational and clean data requirements. In this paper, we propose a novel solution, Chain-of-Scrutiny (CoS), to address these challenges. Backdoor attacks fundamentally create a shortcut from the trigger to the target output, thus lack reasoning support. Accordingly, CoS guides the LLMs to generate detailed reasoning steps for the input, then scrutinizes the reasoning process to ensure consistency with the final answer. Any inconsistency may indicate an attack. CoS only requires black-box access to LLM, offering a practical defense, particularly for API-accessible LLMs. It is user-friendly, enabling users to conduct the defense themselves. Driven by natural language, the entire defense process is transparent to users. We validate the effectiveness of CoS through extensive experiments across various tasks and LLMs. Additionally, experiments results shows CoS proves more beneficial for more powerful LLMs.

Federated Learning (FL) addresses critical issues in machine learning related to data privacy and security, yet suffering from data insufficiency and imbalance under certain circumstances. The emergence of foundation models (FMs) offers potential solutions to the limitations of existing FL frameworks, e.g., by generating synthetic data for model initialization. However, due to the inherent safety concerns of FMs, integrating FMs into FL could introduce new risks, which remains largely unexplored. To address this gap, we conduct the first investigation on the vulnerability of FM integrated FL (FM-FL) under adversarial threats. Based on a unified framework of FM-FL, we introduce a novel attack strategy that exploits safety issues of FM to compromise FL client models. Through extensive experiments with well-known models and benchmark datasets in both image and text domains, we reveal the high susceptibility of the FM-FL to this new threat under various FL configurations. Furthermore, we find that existing FL defense strategies offer limited protection against this novel attack approach. This research highlights the critical need for enhanced security measures in FL in the era of FMs.

Photoplethysmogram (PPG) signals are easily contaminated by motion artifacts in real-world settings, despite their widespread use in Internet-of-Things (IoT) based wearable and smart health devices for cardiovascular health monitoring. This study proposed a lightweight deep neural network, called Tiny-PPG, for accurate and real-time PPG artifact segmentation on IoT edge devices. The model was trained and tested on a public dataset, PPG DaLiA, which featured complex artifacts with diverse lengths and morphologies during various daily activities of 15 subjects using a watch-type device (Empatica E4). The model structure, training method and loss function were specifically designed to balance detection accuracy and speed for real-time PPG artifact detection in resource-constrained embedded devices. To optimize the model size and capability in multi-scale feature representation, the model employed depth-wise separable convolution and atrous spatial pyramid pooling modules, respectively. Additionally, the contrastive loss was also utilized to further optimize the feature embeddings. With additional model pruning, Tiny-PPG achieved state-of-the-art detection accuracy of 87.4% while only having 19,726 model parameters (0.15 megabytes), and was successfully deployed on an STM32 embedded system for real-time PPG artifact detection. Therefore, this study provides an effective solution for resource-constraint IoT smart health devices in PPG artifact detection.

Change detection (CD) is a critical task in studying the dynamics of ecosystems and human activities using multi-temporal remote sensing images. While deep learning has shown promising results in CD tasks, it requires a large number of labeled and paired multi-temporal images to achieve high performance. Pairing and annotating large-scale multi-temporal remote sensing images is both expensive and time-consuming. To make deep learning-based CD techniques more practical and cost-effective, we propose an unsupervised single-temporal CD framework based on intra- and inter-image patch exchange (I3PE). The I3PE framework allows for training deep change detectors on unpaired and unlabeled single-temporal remote sensing images that are readily available in real-world applications. The I3PE framework comprises four steps: 1) intra-image patch exchange method is based on an object-based image analysis method and adaptive clustering algorithm, which generates pseudo-bi-temporal image pairs and corresponding change labels from single-temporal images by exchanging patches within the image; 2) inter-image patch exchange method can generate more types of land-cover changes by exchanging patches between images; 3) a simulation pipeline consisting of several image enhancement methods is proposed to simulate the radiometric difference between pre- and post-event images caused by different imaging conditions in real situations; 4) self-supervised learning based on pseudo-labels is applied to further improve the performance of the change detectors in both unsupervised and semi-supervised cases. Extensive experiments on two large-scale datasets demonstrate that I3PE outperforms representative unsupervised approaches and achieves F1 value improvements of 10.65% and 6.99% to the SOTA method. Moreover, I3PE can improve the performance of the ... (see the original article for full abstract)

Buildings are the basic carrier of social production and human life; roads are the links that interconnect social networks. Building and road information has important application value in the frontier fields of regional coordinated development, disaster prevention, auto-driving, etc. Mapping buildings and roads from very high-resolution (VHR) remote sensing images have become a hot research topic. However, the existing methods often ignore the strong spatial correlation between roads and buildings and extract them in isolation. To fully utilize the complementary advantages between buildings and roads, we propose a building-road collaborative extraction method based on multi-task and cross-scale feature interaction to improve the accuracy of both tasks in a complementary way. A multi-task interaction module is proposed to interact information across tasks and preserve the unique information of each task, which tackle the seesaw phenomenon in multitask learning. By considering the variation in appearance and structure between buildings and roads, a cross-scale interaction module is designed to automatically learn the optimal reception field for different tasks. Compared with many existing methods that train each task individually, the proposed collaborative extraction method can utilize the complementary advantages between buildings and roads by the proposed inter-task and inter-scale feature interactions, and automatically select the optimal reception field for different tasks. Experiments on a wide range of urban and rural scenarios show that the proposed algorithm can achieve building-road extraction with outstanding performance and efficiency.

The efficacy of building footprint segmentation from remotely sensed images has been hindered by model transfer effectiveness. Many existing building segmentation methods were developed upon the encoder-decoder architecture of U-Net, in which the encoder is finetuned from the newly developed backbone networks that are pre-trained on ImageNet. However, the heavy computational burden of the existing decoder designs hampers the successful transfer of these modern encoder networks to remote sensing tasks. Even the widely-adopted deep supervision strategy fails to mitigate these challenges due to its invalid loss in hybrid regions where foreground and background pixels are intermixed. In this paper, we conduct a comprehensive evaluation of existing decoder network designs for building footprint segmentation and propose an efficient framework denoted as BFSeg to enhance learning efficiency and effectiveness. Specifically, a densely-connected coarse-to-fine feature fusion decoder network that facilitates easy and fast feature fusion across scales is proposed. Moreover, considering the invalidity of hybrid regions in the down-sampled ground truth during the deep supervision process, we present a lenient deep supervision and distillation strategy that enables the network to learn proper knowledge from deep supervision. Building upon these advancements, we have developed a new family of building segmentation networks, which consistently surpass prior works with outstanding performance and efficiency across a wide range of newly developed encoder networks. The code will be released on https://github.com/HaonanGuo/BFSeg-Efficient-Building-Footprint-Segmentation-Framework.