Large-scale Data Research Articles

Divide-aggregate heterogeneous hypergraph for large-scale user intention detection

Detecting the intention of users in the information retrieval system is a critical task affecting the user experience. Confronting the challenge of retrieval datasets can be extremely large, complex, and heterogeneous correlated, we propose a framework named DAHH, which utilizes the Divide and Aggregate strategy to incorporate Heterogeneous Hypergraph for large-scale user retrieval intention detection. In the proposed framework, the large-scale retrieval dataset is first divided into multiple subsets with a solvable scale of samples for the model. Then consistent rules are used in all the subsets to construct heterogeneous hypergraphs, and the hypergraph convolution operation is conducted independently on each hypergraph to generate local high-order topologic enhanced vertex representations. Aggregation is applied to correlate all the local hypergraphs at different levels to generate global high-order heterogeneous enhanced representations. In the end, local and global high-order representations are combined to predict user intentions. The proposed framework is capable of dealing with heterogeneous and high-order correlations among the large-scale user retrieval data. We conduct extensive experiments on our collected real-world search engine dataset with millions of samples as well as several widely used public datasets to demonstrate the across-the-board superior performance of DAHH to other state-of-the-art models.

BGS: Accelerate GNN training on multiple GPUs

Emerging Graph Neural Networks (GNNs) have made significant progress in processing graph-structured data, yet existing GNN frameworks face scalability issues when training large-scale graph data using multiple GPUs. Frequent feature data transfers between CPUs and GPUs are a major bottleneck, and current caching schemes have not fully considered the characteristics of multi-GPU environments, leading to inefficient feature extraction. To address these challenges, we propose BGS, an auxiliary framework designed to accelerate GNN training from a data perspective in multi-GPU environments. Firstly, we introduce a novel training set partition algorithm, assigning independent training subsets to each GPU to enhance the spatial locality of node access, thus optimizing the efficiency of the feature caching strategy. Secondly, considering that GPUs can communicate at high speeds via NVLink connections, we designed a feature caching placement strategy suitable for multi-GPU environments. This strategy aims to improve the overall hit rate by setting reasonable redundant caches on each GPU. Evaluations on two representative GNN models, GCN and GraphSAGE, show that BGS significantly improves the hit rate of feature caching strategies in multi-GPU environments and substantially reduces the time overhead of data loading, achieving a performance improvement of 1.5 to 6.2 times compared to the baseline.

Causal relationships between serum albumin, neuroticism and suicidal ideation in depressed patients: A Mendelian randomization study

Although serum albumin and neuroticism have revealed a strong association with suicidal ideation in individuals with depression, the causal relationship between them is uncertain. This study analyzed the causal association of serum albumin, neuroticism and suicidal ideation using large-scale GWAS data and Univariable Mendelian Randomization (UVMR) methods. The Multivariable MR (MVMR) analysis was used to explore the causal pathways. UVMR analysis revealed that genetically determined serum albumin is causally associated with neuroticism (β = −0.006 S.D.; 95 % CI: 0.009, −0.002; p = 0.003) and suicidal ideation (β = 0.009 S.D.; 95 % CI: 0.001, 0.016; p = 0.037); and that neuroticism mediates 100 % of the causal association between serum albumin and suicidal ideation in individuals with depression. These findings suggest genetic evidence for the causal effect of serum albumin on suicidal ideation in depressed patients and the significant mediation effect of neuroticism on this causal association. This study proves the protective role of serum albumin for neuroticism and the riskiness of personality traits for suicidal ideation in individuals with depression.

Associations between behavioral and self-reported impulsivity, brain structure, and genetic influences in middle childhood

Impulsivity undergoes a normative developmental trajectory from childhood to adulthood and is thought to be driven by maturation of brain structure. However, few large-scale studies have assessed associations between impulsivity, brain structure, and genetic susceptibility in children. In 9112 children ages 9–10 from the ABCD study, we explored relationships among impulsivity (UPPS-P impulsive behavior scale; delay discounting), brain structure (cortical thickness (CT), cortical volume (CV), and cortical area (CA)), and polygenic scores for externalizing behavior (PGSEXT). Both higher UPPS-P total scores and more severe delay-discounting had widespread, low-magnitude associations with smaller CA in frontal and temporal regions. No associations were seen between impulsivity and CV or CT. Additionally, higher PGSEXT was associated with both higher UPPS-P scores and with smaller CA and CV in frontal and temporal regions, but in non-overlapping cortical regions, underscoring the complex interplay between genetics and brain structure in influencing impulsivity. These findings indicate that, within large-scale population data, CA is significantly yet weakly associated with each of these impulsivity measures and with polygenic risk for externalizing behaviors, but in distinct brain regions. Future work should longitudinally assess these associations through adolescence, and examine associated functional outcomes, such as future substance use and psychopathology.

Arctic mixed-phase clouds simulated by the WRF model: Comparisons with ACLOUD radar and in situ airborne observations and sensitivity of microphysics properties

The microphysical properties of two mixed-phase clouds (MPCs), one over sea ice, and another over open ocean, are investigated using in situ measuring probes and a cloud radar installed on-board two aircraft during the Arctic CLoud Observations Using airborne measurements during polar Day (ACLOUD) field campaign, west of Svalbard on 17 June 2017. Comparisons between observations and different configurations of the Weather and Research Forecasting (WRF) model are conducted and reveal discrepancies in the vertical profile of the ice crystal concentrations for both MPCs due to slightly higher temperatures (by approx. 1–2 °C) in WRF than observed, affecting the ice formation. A comprehensive sensitivity analysis is carried out to address the impact of assumptions used for the microphysics (MP) and the atmospheric boundary layer (ABL) schemes on the simulated vertical properties of the MPC systems. The role of the vertical resolution of the large scale data used for model initialisation is also investigated. Results show that the model performance improves for large-scale initialisation with more numerous levels close to the surface. Moreover, modifying the ABL or MP scheme influences the altitudes of the cloud top and base as well as the liquid and ice water amounts. Our study also highlights how an increase in the cloud condensation nuclei and droplet number concentrations can impact the vertical profiles of the liquid and ice water amounts.

Identifying the association between depression and constipation: An observational study and Mendelian randomization analysis

BackgroundBoth depression and constipation are universal disorders that seriously affect quality of life. But the phenotypic relationship and causality between depression and constipation are still unclear. MethodsWe first assessed phenotypic relationships by logistic regression analysis using large-scale data extracted from the National Health and Nutrition Examination Survey (N = 11,585). We then evaluated causality by bidirectional two-sample mendelian randomization (MR) analysis using Genome-wide association study (GWAS) data (depression: N = 807,553; constipation: N = 377,277). To investigate whether depression severity affects the causal relationship between depression and constipation, we conducted a further MR study on GWAS data of major depression (N = 480,359). ResultsAbout 11.31 % of the participants in the constipation group suffered from depression, which was significantly higher than the normal bowel group (6.09 %). The observational study showed a positive correlation between depression and constipation (OR = 1.968, 95%CI = 1.530–2.532). Besides, the risk of constipation was higher in participants with severe depression (OR = 2.294, 95%CI = 1.538–3.422) than in participants with mild depression (OR = 1.549, 95%CI = 1.242–1.932). Bidirectional MR analysis revealed an obviously causal effect of depression on constipation, but no causal effect of constipation on depression. In addition, the MR analysis also revealed a causal relationship between major depression and constipation. LimitationThe exact mechanism by which depression affects constipation is still unclear. ConclusionThis study reveals a positive correlation between depression and constipation and the causal effect of depression on constipation. Clinicians should keep the risk of constipation in mind when treating patients with depression.

AdaDFKD: Exploring adaptive inter-sample relationship in data-free knowledge distillation

In scenarios like privacy protection or large-scale data transmission, data-free knowledge distillation (DFKD) methods are proposed to learn Knowledge Distillation (KD) when data is not accessible. They generate pseudo samples by extracting the knowledge from teacher model, and utilize above pseudo samples for KD. The challenge in previous DFKD methods lies in the static nature of their target distributions and they focus on learning the instance-level distributions, causing its reliance on the pretrained teacher model. To address above concerns, our study introduces a novel DFKD approach known as AdaDFKD, designed to establish and utilize relationships among pseudo samples, which is adaptive to the student model, and finally effectively mitigates the aforementioned risk. We achieve this by generating from “easy-to-discriminate” samples to “hard-to-discriminate” samples as human does. We design a relationship refinement module (R2M) to optimize the generation process, wherein we learn a progressive conditional distribution of negative samples and maximize the log-likelihood of inter-sample similarity of pseudosamples. Theoretically, we discover that such design of AdaDFKD both minimize the divergence and maximize the mutual information between the distribution of teacher and student models. Experimental results demonstrate the superiority of our approach over state-of-the-art (SOTA) DFKD methods across various benchmarks, teacher-student pairs, and evaluation metrics, as well as well robustness and fast convergence.

AoI-Guaranteed Incentive Mechanism for Mobile Crowdsensing With Freshness Concerns

With the explosive spread of smart mobile devices, Mobile CrowdSensing (MCS) has been becoming a promising paradigm, by which a platform can coordinate a group of workers to complete large-scale data collection tasks using their mobile devices. In this paper, we investigate the incentive mechanism design in MCS systems, taking the freshness of collected data and social benefits into consideration. First, the Age of Information (AoI) metric is introduced to measure the freshness of data. Then, we model the incentive mechanism design with AoI guarantees as an incomplete information two-stage Stackelberg game with multiple constraints. Next, we consider the scenario that all participants share the public utility function parameters of the Stackelberg game. By deriving the optimal remuneration paid by the platform and the optimal data update frequency for each worker, and proving the existence of a unique Stackelberg equilibrium, we propose an AoI-guaranteed Incentive Mechanism (AIM) that enables the platform and all workers to maximize their utilities simultaneously. Furthermore, we extend AIM to a general scenario where each participant has no prior knowledge of the utility function parameters of the game. By resorting to the Deep Reinforcement Learning (DRL) technique and modeling the two-stage Stackelberg game as a Markov decision process, we propose a DRL-based Incentive Mechanism (DIM) with AoI guarantees, which makes each participant effectively seek its optimal strategy through trial and error. Meanwhile, the system can guarantee that the AoI values of all data uploaded to the platform are not larger than a given threshold. Finally, numerical experiments on real-world traces are conducted to validate the efficacy and efficiency of AIM and DIM.

Evolutionary trends of respiratory syncytial viruses: Insights from large-scale surveillance and molecular dynamics of G glycoprotein

Human respiratory syncytial virus (RSV) is an underlying cause of lower respiratory illnesses in children, elderly and immunocompromised adults. RSV contains multiple structural and non-structural proteins with two major glycoproteins that control the initial phase of infection, fusion glycoprotein and the attachment (G) glycoprotein. G protein attaches to the ciliated cells of airways initiating the infection. The hypervariable G protein plays a vital role in evolution of RSV strains. We employed multiple bioinformatics tools on systematically accessed large-scale data to evaluate mutations, evolutionary history, and phylodynamics of RSV. Mutational analysis of central conserved region (CCR) on G protein-coding sequences between 163 and 189 positions revealed frequent mutations at site 178 in human RSV (hRSV) A while arginine to glutamine substitutions at site 180 positions in hRSV B, remained prevalent from 2009 to 2014. Phylogenetic analysis indicates multiple signature mutations within G protein responsible for diversification of clades. The USA and China have highest number of surveillance records, followed by Kenya. Markov Chain Monte Carlo Bayesian skyline plot revealed that RSV A evolved steadily from 1990 to 2000, and rapidly between 2003 and 2005. Evolution of RSV B continued from 2003 to 2022, with a high evolution stage from 2016 to 2020. Throughout evolution, cysteine residues maintained their strict conserved states while CCR has an entropy value of 0.0039(±0.0005). This study concludes the notion that RSV G glycoprotein is continuously evolving while the CCR region of G protein maintains its conserved state providing an opportunity for CCR-specific monoclonal antibodys (mAbs) and inhibitors as potential candidates for immunoprophylaxis.

GA-GGD: Improving semantic discriminability in graph contrastive learning via Generative Adversarial Network

Graph contrastive learning has garnered considerable research interest due to its ability to effectively embed graph data without manual labels. Among them, methods based on Deep Graph Infomax (DGI) have been widely studied and favored in the industry because of their fast training speed, applicability to large-scale data. DGI-based methods usually obtain a noise graph through node shuffling. The proxy task of these methods encourages the encoder to distinguish whether the nodes come from the original graph or the noise graph, thereby maximizing the mutual information between the node representation and the graph it belongs to, while also maximizing the Jenson-Shannon divergence between the nodes of original graph and noise graph. However, we argue that these approaches only enable the encoder to differentiate between semantically meaningful graphs and noise graphs, but not to effectively identify different semantic graphs. This leads to the inability of the encoder to effectively embed information between different semantics, significantly reducing the robustness and affecting the performance of downstream tasks. In addition, this training mode makes the model more sensitive to attacks. To improve their semantic discriminability, we take advantage of the natural ability of generative adversarial networks to generate semantic data, proposing a method called Generative Adversarial Graph Group Discrimination (GA-GGD). Specifically, it consists of a graph group discriminator and a semantic attack generator. The discriminator aims to encode the graph and identify whether nodes originate from the original graph. The goal of the generator is to use random features and graph structure to find vulnerabilities of discriminator and generate node representations with similar but wrong semantic to confuse the discriminator. GA-GGD can improve the model’s semantic information embedding without significantly increasing computational overhead and memory occupancy. We test the effectiveness of the proposed model on commonly used data sets and large-scale datasets, as well as in various downstream tasks such as classification, clustering, and adversarial attacks defence. A wealth of experimental results confirm the efficacy of the proposed model.

A memory-friendly class-incremental learning method for hand gesture recognition using HD-sEMG

Hand gesture recognition (HGR) plays a vital role in human-computer interaction. The integration of high-density surface electromyography (HD-sEMG) and deep neural networks (DNNs) has significantly improved the robustness and accuracy of HGR systems. These methods are typically effective for a fixed set of trained gestures. However, the need for new gesture classes over time poses a challenge. Introducing new classes to DNNs can lead to a substantial decrease in accuracy for previously learned tasks, a phenomenon known as "catastrophic forgetting," especially when the training data for earlier tasks is not retained and retrained. This issue is exacerbated in embedded devices with limited storage, which struggle to store the large-scale data of HD-sEMG. Class-incremental learning (CIL) is an effective method to reduce catastrophic forgetting. However, existing CIL methods for HGR rarely focus on reducing memory load. To address this, we propose a memory-friendly CIL method for HGR using HD-sEMG. Our approach includes a lightweight convolutional neural network, named SeparaNet, for feature representation learning, coupled with a nearest-mean-of-exemplars classifier for classification. We introduce a priority exemplar selection algorithm inspired by the herding effect to maintain a manageable set of exemplars during training. Furthermore, a task-equal-weight exemplar sampling strategy is proposed to effectively reduce memory load while preserving high recognition performance. Experimental results on two datasets demonstrate that our method significantly reduces the number of retained exemplars to only a quarter of that required by other CIL methods, accounting for less than 5% of the total samples, while still achieving comparable average accuracy.

Cascade Residual Multiscale Convolution and Mamba-Structured UNet for Advanced Brain Tumor Image Segmentation

In brain imaging segmentation, precise tumor delineation is crucial for diagnosis and treatment planning. Traditional approaches include convolutional neural networks (CNNs), which struggle with processing sequential data, and transformer models that face limitations in maintaining computational efficiency with large-scale data. This study introduces MambaBTS: a model that synergizes the strengths of CNNs and transformers, is inspired by the Mamba architecture, and integrates cascade residual multi-scale convolutional kernels. The model employs a mixed loss function that blends dice loss with cross-entropy to refine segmentation accuracy effectively. This novel approach reduces computational complexity, enhances the receptive field, and demonstrates superior performance for accurately segmenting brain tumors in MRI images. Experiments on the MICCAI BraTS 2019 dataset show that MambaBTS achieves dice coefficients of 0.8450 for the whole tumor (WT), 0.8606 for the tumor core (TC), and 0.7796 for the enhancing tumor (ET) and outperforms existing models in terms of accuracy, computational efficiency, and parameter efficiency. These results underscore the model’s potential to offer a balanced, efficient, and effective segmentation method, overcoming the constraints of existing models and promising significant improvements in clinical diagnostics and planning.

Characterization of Aerosol and CO2 Co-Emissions around Power Plants through Satellite-Based Synergistic Observations

The tracking of carbon and aerosol co-emissions is essential for environmental management. Satellite-based atmospheric synoptic observation networks provide large-scale and multifaceted data to help resolve emission behaviors. This study employs a comprehensive analysis of atmospheric dynamics, combustion byproducts, and aerosol characteristics around power plants. Strong correlations between Aerosol Optical Depth (AOD) at 500 nm and the column-averaged dry-air mole fraction of carbon dioxide (XCO2) were observed, revealing synchronous peaks in their emission patterns. The investigation into combustion completeness utilized metrics such as the ratio of carbon monoxide (CO)/XCO2 and Black Carbon Extinction (BCEXT)/Total Aerosol Extinction (TOTEXT). Discrepancies in these ratios across cases suggest variations in combustion efficiency and aerosol characteristics. Nitrogen dioxide (NO2) distributions closely mirrored XCO2, indicating consistent emission patterns, while variations in sulfur dioxide (SO2) distributions implied differences in sulfide content in the coal used. The influence of coal composition on AOD/XCO2 ratios was evident, with sulfide content contributing to variations besides combustion efficiency. This multifactorial analysis underscores the complex interplay of combustion completeness, aerosol composition, and coal components in shaping the air quality around power stations. The findings highlight the need for a nuanced understanding of these factors for effective air quality management.

Traffic trajectory generation via conditional Generative Adversarial Networks for transportation Metaverse

The transportation Metaverse, by integrating real and virtual vehicular networks, brings significant benefits to the development of smart cities. However, the difficulty and high cost of conducting large-scale traffic and driving simulations in the transportation Metaverse via realistic data collection and fusion from the physical world directly result in the lack of spatial–temporal traffic data and privacy concerns, which significantly hinder the development of the transportation Metaverse. Hence, in these situations, it becomes essential to produce high-quality, large-scale trajectory data to support relevant applications. However, the deficiency of data-driven method is the heavy dependence on high-quality historical data. Prior knowledge can help reduce learning difficulties and overfitting problems with small amounts of data. In our study, we use a hybrid framework, the Travel Demand Conditioning Generative Adversarial Network (TD-GAN), which combines data-driven and knowledge-driven approaches to address the issue of traffic trajectory generation. First, we employ a conditional mechanism to incorporate prior knowledge of travel demand to reduce learning difficulties. Second, non-standard convolutional module and multi-headed self-attention module are developed to capture spatial–temporal correlations. The experimental results show that our model outperforms the baseline models.

Individual and community empowerment improve resource users' perceptions of community-based conservation effectiveness in Kenya and Tanzania.

Community-based conservation has been increasingly recognized as critical to achieve both conservation and socio-economic development goals worldwide. However, the long-term sustainability of community-based conservation programs is dependent on a broadly shared perception among community members that management actions are achieving their stated goals. Thus, understanding the underlying factors driving differences in perceptions of management effectiveness can help managers prioritize the processes and outcomes most valued by resource users and thereby promote sustained support for conservation efforts. Here, we utilize large-scale interview survey data and machine learning to identify the factors most strongly associated with differences in perceived management effectiveness between resource users engaged in marine community-based conservation programs in Kenya and Tanzania. Perceptions of management effectiveness were generally favorable in both countries, and the most important predictors of positive perceptions were associated with community and individual empowerment in resource management and use, but within disparate focal domains. Improved perceptions of management effectiveness in Kenya were closely related to increases in women's empowerment in community-based conservation programs, while inclusionary and transparent governance structures were the most important factors driving improved perceptions in Tanzania. Additionally, the strongest predictors of differences between individuals in both countries often interacted synergistically to produce even higher rates of perceived effectiveness. These findings can help future initiatives in the region tailor management to match community-level priorities and emphasize the need for community-based conservation programs to understand local context to ensure that metrics of "success" are aligned with the needs and desires of local resource users.

A Comprehensive Survey of Artificial Intelligence and Machine Learning Application in Healthcare

The ongoing technological transformation, driven by rapid advancements in machine learning (ML), holds the potential to profoundly impact our lives and redefine the essence of humanity. Recent years have witnessed significant progress in ML, fuelled by the widespread adoption of artificial intelligence (AI) due to machines' capacity for large-scale data processing and management. The intersection of ML and healthcare encompasses both computer science along with medical science. Within medical science, ML techniques have enabled the examination of complex medical data, marking a significant advancement. In the healthcare industry, ML functions as the cognitive and knowledge counterpart of healthcare professionals. This paper explores the benefits of ML-based solutions and their applications in healthcare, emphasizing their role in automating data analysis for patients' health records along with making predictions per the extracted insights. The integration of ML in healthcare promises transformative outcomes, shaping the future of medical diagnostics and decision-making.

Using personal tracking software data to aid the impact of illegal trails in national parks

ABSTRACT The use of personal tracking devices for recreational use and for safety when in the outdoors has grown exponentially in recent years. The large-scale data collected through these platforms has enabled our team to gain a greater understanding of the way in which people encroach onto National Parks causing environmental damage. The project used freely accessible data on the access of illegal trails primarily from motor bike and mountain bike use in the East Otway Range of Victoria. This region was an excellent case study as it is a mosaic of state and national parks and the signage and restrictive fences place by national parks have been aligned with the illegal trail formation. The data mapping has been coupled with measurement of the impact on the environment including the number and scale of the trails and the type of environment they are impacting including fauna and flora of national significance. The data has shown some useful information around the manner in which illegal trails are identified and concealed by users. Further, the research has uncovered ways in which trails have been concealed from authorities.

Generalized Weak Supervision for Neural Information Retrieval

Neural ranking models (NRMs) have demonstrated effective performance in several information retrieval (IR) tasks. However, training NRMs often requires large-scale training data, which is difficult and expensive to obtain. To address this issue, one can train NRMs via weak supervision, where a large dataset is automatically generated using an existing ranking model (called the weak labeler) for training NRMs. Weakly supervised NRMs can generalize from the observed data and significantly outperform the weak labeler. This paper generalizes this idea through an iterative re-labeling process, demonstrating that weakly supervised models can iteratively play the role of weak labeler and significantly improve ranking performance without using manually labeled data. The proposed Generalized Weak Supervision (GWS) solution is generic and orthogonal to the ranking model architecture. This paper offers four implementations of GWS: self-labeling, cross-labeling, joint cross- and self-labeling, and greedy multi-labeling. GWS also benefits from a query importance weighting mechanism based on query performance prediction methods to reduce noise in the generated training data. We further draw a theoretical connection between self-labeling and Expectation-Maximization. Our experiments on four retrieval benchmarks suggest that our implementations of GWS lead to substantial improvements compared to weak supervision if the weak labeler is sufficiently reliable.

Development and implementation of ion cyclotron emission diagnostic system on J-TEXT tokamak

An ion cyclotron emission (ICE) diagnostic based on a magnetic probe has been designed and installed on the J-TEXT tokamak. This ICE diagnostic is used to study the high-frequency magnetic fluctuations driven by energetic particles. The signals detected by the ICE probe are divided into two equal power signals, which enter the two systems respectively. The first system converts high-frequency signals into power signals in different frequency bands using bandpass filters and power detectors. This reduces the amount of data to be processed and allows the data to be uploaded in time after a discharge. Thus, the approximate time resolution, frequency range (10 MHz-150 MHz), and power of the high-frequency signals can be quickly judged. The second system uses low-pass filters and an oscilloscope to collect data by Nyquist sampling law, so that the frequency spectrum of signals with frequencies from 7 kHz to 200 MHz can be analyzed. Therefore, if significant signals are detected in the first system on the data server, corresponding stored data can be extracted from the second system for large-scale data processing. This structure enhances the efficiency of data storage and analysis. The high-frequency magnetic fluctuations driven by fast electrons or runaway electrons under ohmic heating have been measured on J-TEXT, which verified the feasibility of the system.

Invisible Black-Box Backdoor Attack against Deep Cross-Modal Hashing Retrieval

Deep cross-modal hashing has promoted the field of multi-modal retrieval due to its excellent efficiency and storage, but its vulnerability to backdoor attacks is rarely studied. Notably, current deep cross-modal hashing methods inevitably require large-scale training data, resulting in poisoned samples with imperceptible triggers that can easily be camouflaged into the training data to bury backdoors in the victim model. Nevertheless, existing backdoor attacks focus on the uni-modal vision domain, while the multi-modal gap and hash quantization weaken their attack performance. In addressing the aforementioned challenges, we undertake an invisible black-box backdoor attack against deep cross-modal hashing retrieval in this article. To the best of our knowledge, this is the first attempt in this research field. Specifically, we develop a flexible trigger generator to generate the attacker’s specified triggers, which learns the sample semantics of the non-poisoned modality to bridge the cross-modal attack gap. Then, we devise an input-aware injection network, which embeds the generated triggers into benign samples in the form of sample-specific stealth and realizes cross-modal semantic interaction between triggers and poisoned samples. Owing to the knowledge-agnostic of victim models, we enable any cross-modal hashing knockoff to facilitate the black-box backdoor attack and alleviate the attack weakening of hash quantization. Moreover, we propose a confusing perturbation and mask strategy to induce the high-performance victim models to focus on imperceptible triggers in poisoned samples. Extensive experiments on benchmark datasets demonstrate that our method has a state-of-the-art attack performance against deep cross-modal hashing retrieval. Besides, we investigate the influences of transferable attacks, few-shot poisoning, multi-modal poisoning, perceptibility, and potential defenses on backdoor attacks. Our codes and datasets are available at https://github.com/tswang0116/IB3A.