Latest ArticlesPre-trained models have mitigated the challenges posed by extensive training data and computational resources, and also give birth to the new paradigm of model development and application, which we refer to as model supply chain. In this framework, a pre-trained model is uploaded by its publisher and subsequently transferred, compressed, and deployed by secondary developers to meet various application needs. This emerging model supply chain introduces additional stages and multiple elements, inevitably leading to security concerns and privacy risks. Despite the widespread adoption of model supply chains, there is currently a lack of systematic review of security threats in them. To address this research gap, we provide a comprehensive overview of the deep learning model supply chain, introducing its concept and fundamental structure. We conduct an in-depth analysis of vulnerabilities at various stages of the model’s lifecycle, including design, development, deployment, and usage. Furthermore, we compare and summarize prevalent attack methods, alongside introducing corresponding security protection strategies. To assist readers in effectively utilizing pre-trained models, we review and compare publicly available model repositories. Finally, we discuss potential future research avenues in areas such as security checks, real-time detection, and problem tracing. This review aims to offer insights for safer and more reliable development and use of pre-training models. For the benefit of ongoing research, related papers and open-source codes of the methods discussed are accessible at website: https://github.com/Dipsy0830/DNN-supply-chain-survey.
Software systems play an indispensable role across various industries, handling large-scale and high-density data. However, the numerous defects within these systems have troubled developers for a long time, constantly threatening the security of data elements. Automated program repair (APR) technology aims to assist developers in automatically fixing defects in code during software development process, thereby saving costs in software system development and maintenance, enhancing the confidentiality, availability, and integrity of data elements within software systems. With the development of large language model (LLM) technology, many powerful code large language models have emerged. These models have demonstrated strong repair capabilities in the APR field, while also addressing shortcomings of traditional approaches in code comprehension and patch generation capabilities, further elevating the level of program repair tools. We thoroughly survey high-quality papers related to APR in recent years, summarizing the latest developments in the field. We then systematically categorize two types of LLM-based APR techniques: cloze style and neural machine translation style. We also conduct an in-depth comparison from various perspectives such as model usage, model size, types of defects repaired, programming languages involved, and the pros and cons of repair approaches. Additionally, we discuss the widely adopted APR datasets and metrics, and outline existing empirical studies. Finally, we summarize current challenges in the APR field along with future research directions.
Transient execution attacks (TEAs) exploit processor optimizations to bypass security checks and exfiltrate sensitive information through covert channels. Among them, Meltdown and Spectre attacks have become prominent, affecting mainstream commercial processors such as Intel, ARM, and AMD. Despite the defensive measures implemented by processor manufacturers, variants of these attacks continue to be discovered and disclosed by researchers. To improve the understanding of TEAs and deploy robust defenses, this paper comprehensively analyzes TEAs under various covert channels. Initially, the common characteristics of TEAs are extracted, and a novel model for TEAs is systematically constructed. Subsequently, we summarize the various types of covert channels involved in existing research, classify the TEAs into three types: Meltdown type attacks driven by out-of-order execution (OoOE), Spectre type attacks driven by branch misprediction, and microarchitecture data sampling (MDS) type attacks driven by data misprediction, and delineate the key aspects and relationships of each type of attack. Notably, this paper systematically compiles and categorizes MDS type attacks for the first time. Then, the capabilities of each attack variant are meticulously analyzed and evaluated from three dimensions: covert channel, attack applicable scenarios, and microarchitecture immunity status, which aids security researchers in developing new and more destructive attack types based on the deficiencies of the existing attack-related research. Finally, combined with the above-mentioned comprehensive and in-depth analysis, and the summary of processor microarchitecture and covert channels, this paper anticipates the future trajectory of TEAs research, hoping to provide strong support for subsequent research work.
Given the risk of adversarial attacks on tracking models and the lack of relevant adversarial detection methods, this paper addresses the problem from the perspective of frequency domain. Combined with the visual invisible property of perturbation noise, we first theoretically prove that perturbation noise mainly exists in the mid-to-high frequency bands of images. Then we quantitatively analyze that the low-frequency components of the video sequence contribute the most to tracking performance and are least affected by adversarial attacks. Finally, based on the above theoretical proof and qualitative analysis, we propose a detection framework based on the tracking performance difference of frequency bands, in which the frequency domain decomposition module for extracting the low-frequency components of the video sequence. The target tracker and its mirror tracker with the same structure and parameters respectively take the full-frequency and low-frequency components of the video sequence as input. The discriminator module determines whether the input video sequence is an adversarial input by comparing the output differences of the two trackers. This detection framework uses a tracker as a carrier and does not require adversarial training. It can achieve adversarial detection only by comparing the tracking performance difference across different frequency bands. Extensive experimental results show that the detection framework can not only effectively detect current mainstream adversarial attacks, such as CSA, TTP, and Spark with a detection precision of 97.55%, but also has little negative impact on the original tracking performance of the tracker. In addition, this framework is generalizable and can be flexibly integrated into multiple trackers, such as SiamRPNpp, SiamMask, SiamCAR, and SiamBAN.
Deep learning-based object detection algorithms have been widely applied, while recent research indicates that these algorithms are vulnerable to adversarial attacks, causing detectors to either misidentify or miss the target. Nonetheless, research focusing on the transferability of adversarial attacks in autonomous driving is limited, and few studies address the stealthiness of such attacks in this scenario. To address these limitations in current research, an algorithmic module to enhance attack transferability is designed by drawing an analogy between optimizing adversarial examples and the training process of machine learning models. Additionally, through employing style transfer techniques and neural rendering, a transferable and stealthy attack (TSA) method is proposed and implemented. Specifically, the adversarial examples are first repeatedly stitched together and combined with masks to generate the final texture, which is then applied to the entire vehicle surface. To simulate real-world conditions, a physical transformation function is used to embed the rendered camouflaged vehicle into realistic scenes. Finally, the adversarial examples are optimized using a designed loss function. Simulation experiments demonstrate that the TSA method surpasses existing methods in attack transferability and exhibits a certain level of stealthiness in appearance. Furthermore, physical domain experiments validate that the TSA method maintains effective attack performance in real-world scenarios.
Leveraging a sliding window strategy, this study presents an innovative retrieval-augmented generation system aimed at enhancing the factual accuracy and reliability of outputs from large language models (LLMs). By applying a sliding window mechanism during the indexing phase, the project effectively addresses the limitations of fixed context window sizes and static retrieval methods. Three specific sliding window strategies are proposed to efficiently process and segment texts, such as fixed window size and fixed step length split (FFS), dynamic window size and fixed step length split (DFS), and dynamic window size and dynamic step length split (DDS). To further enhance retrieval accuracy and relevance, the project employs multiple advanced query techniques, including query expansion and reformulation. Rigorous experimental evaluations are conducted using the state-of-the-art LLaMA-3 model across multiple diverse datasets, encompassing both general knowledge and domain-specific corpora. Results demonstrate optimal performance with a carefully calibrated block size of 1 024 tokens and a step size of 3, significantly improving F1 score across various tasks. This configuration highlights the critical importance of balancing document segment length and sliding window step size to maximize information retention and retrieval efficacy.The sliding window strategy effectively preserves contextual information, reduces information loss, and exhibits adaptability across different datasets and query types.
Sequential recommendation is centered on mining users’ preferences and behavior patterns from their interaction sequences. Existing work has recognized the inadequacy of single-modal interaction data, and has utilized a large amount of multi-modal data, including item reviews, homepage images, and other sources, to complement interaction data and improve recommendation performance. However, these multi-modal data are often interspersed with unavoidable noise that may limit the exploration of personalized user preferences. While suppressing inter-modal inconsistent information can reduce noise interference, it is almost impossible to completely eliminate noise from user-generated multi-modal content. To address the above challenges, we propose a large language model-based trusted multi-modal recommendation (Large-TR) algorithm, which aims to provide the trustworthy recommendation in noisy multi-modal data scenarios. Specifically, the algorithm relies on the excellent natural language understanding capability of the large language model, which efficiently filters the noise in multi-modal data and achieves more accurate and detailed modelling of user preferences. Additionally, we design a trustworthy decision mechanism that dynamically evaluates the uncertainty of recommendation results and ensures the usability of recommendation results in high-risk scenarios. Experimental results on four widely used public datasets show that the algorithm proposed in this paper has better performance compared with other baseline algorithms. Our source code is available at https://github.com/ hhbray/Large-TR.
According to statistics, the number of people suffering from cardiovascular diseases in China is about 330 million, and the number of deaths caused by cardiovascular diseases accounts for 40% of the total number of deaths each year. Under this circumstance, the development of heart disease assisted diagnosis systems is particularly important, but its development is limited by the lack of a large amount of electrocardiogram (ECG) clinical data that don’t contain patient privacy information and need to be annotated by medical experts. As an emerging discipline, quantum computing can explore larger and more complex state spaces by utilizing quantum superposition and entanglement properties, which is beneficial for generating high-quality and diverse electrocardiogram data similar to real clinical data. Therefore, we propose an electrocardiogram generative information system based on quantum generative adversarial networks, abbreviated as ECG-QGAN. The quantum generative adversarial network consists of a quantum bidirectional gated recurrent unit (QBiGRU) and a quantum convolutional neural network (QCNN). The system utilizes the entanglement property of quantum to improve the generative capability to produce ECG data that are consistent with the existing clinical data so that the heartbeat characteristics of cardiac patients can be preserved. The generator and discriminator of this system use QBiGRU and QCNN, respectively. The variational quantum circuit (VQC) designed based on matrix product state (MPS) and tree tensor network (TTN) is adopted, which enables the system to capture ECG data information more efficiently and generate qualified ECG data with fewer quantum resources. In addition, the system applies quantum Dropout technology to avoid overfitting issues during the training process. Finally, the experimental results show that the ECGs generated by ECG-QGAN have a higher average classification accuracy compared with other models for generating ECGs. It is also friendly to the current noise intermediate scale quantum (NISQ) computers in terms of the number of quantum bits and circuit depth.
With the explosive growth of scientific literature and the continuous deepening of research fields, researchers face significant information processing challenges when attempting to formulate novel scientific hypotheses. Although large language models (LLMs) possess considerable potential for data processing and knowledge integration, they remain limited in their ability to generate original and insightful scientific hypotheses. Existing research predominantly emphasizes utilizing LLMs to expedite and refine established theories and technologies, often overlooking the initial stages of scientific inquiry where novel hypotheses are proposed and new theories are developed—a stage vital to scientific advancement. This study, grounded in the principles of divergent and convergent thinking from the theory of structured intelligence, proposes an innovative human-in-the-loop multi-agent framework (HILMA) for the reliable generation of scientific hypotheses. HILMA framework incorporates a real-time, systematic knowledge retrieval enhancement mechanism, dynamically integrating the latest research advancements to construct citation network subgraphs, providing LLMs with comprehensive and up-to-date scientific knowledge surveys. Additionally, the framework enhances hypothesis generation through a multi-agent argumentation approach that simulates the scientific peer review process, while also leveraging the intuition and expertise of human experts to further refine and diversify the generated hypotheses. A series of human-machine evaluations has shown that this method demonstrates significant advantages over existing baselines in generating high-quality scientific hypotheses and holds promise as a key facilitator for driving technological innovation.
With the global population aging and lifestyle changing, the management and treatment of chronic diseases become increasingly important. Chronic diseases include cardiovascular diseases, diabetes, chronic respiratory diseases, etc. They require long-term or even lifelong health management, the core of which is to design and implement long-term health plans, including balanced dieting, appropriate exercising, regular inspection, and medication management. In recent years, large language models have made progress in the medical field but haven’t focused on chronic disease health management. Therefore, they lack understanding of Chinese dietary habits and culture. These medical large language models also have limited capabilities in handling numerical information. To address these issues, we construct a chronic disease health management information system based on large language model. By integrating foundational knowledge of chronic diseases, health management guidelines, and actual health management plans as domain data, we train QingTing large language model as the core of the system for effectively answering health-related questions. Additionally, the system introduces a tool enhancement strategy, improving QingTing’s ability to handle numerical information in health data by invoking tools. The system also adopts a retrieval-augmented generation technology based on uncertain knowledge graph to enhance the accuracy and reliability of QingTing. Experiments on the chronic disease health management information system based on a large language model demonstrate that QingTing significantly outperforms other baseline large language models in health management dialogues, and verifies the effectiveness of the designed tool enhancement and retrieval-augmented methods.