Surgical scene understanding is a vital intelligent technique in robot-assisted surgery, including surgical instrument detection, segmentation, and instrument–tissue interaction detection. Existing methods typically address these tasks in isolation, neglecting the intrinsic correlations among them. In this work, we innovatively propose a unified multitask framework named UniSurg, being the first to jointly address these three critical aspects of surgical scene understanding, thereby providing the robot with multidimensional perceptual capabilities. By exploring the inter-task correlations and reusing shared features, UniSurg has been demonstrated to significantly enhance the scene analysis performance. To address pose variability of the instruments under the constrained field of view in laparoscopic surgery, we design an Attention Enhanced Conditional Convolution (AEC-Conv) that dynamically adjusts kernels based on pose-specific features for improved adaptability. To further enhance interaction detection, we propose the Temporal Difference Enhancement module (TDE), which captures motion cues by amplifying inter-frame differences, and the Pyramid Global Feature Enhancement module (PGFE), which leverages graph-based hierarchical context to model global relational dependencies. Experiments on the Endovis2018 dataset and a clinical multitask dataset MILVis demonstrate the superior multitask performance of UniSurg.

Federated learning, as a distributed machine learning framework, enables participants to collaboratively train models by uploading only local gradients instead of exchanging local data. However, the malicious server might infer the participants' private data from the uploaded gradients or return incorrect aggregated results to participants. To tackle the above issues, numerous privacy-preserving and verifiable federated learning schemes have been developed. However, only a few of these schemes address collusion-resistant verification, and none considers the potential disputes between participants and server. In this paper, we put forward a privacy-preserving federated learning scheme with collusion-resistant verification and fair arbitration (PFLVA). In PFLVA, a novel verification method is designed, in which non-colluding participants can verify the correctness of aggregated gradient, even if up to$N-1$participants collude with the server, where$N$denotes the total number of online participants. We propose an efficient gradient encryption method to ensure participants' privacy while substantially reducing the computational overhead. We introduce a smart contract to locate the compromised entity when disputes arise and to achieve fair arbitration. Additionally, PFLVA allows participants to go offline without incurring additional computational or communication overheads for the online participants. We provide a comprehensive security analysis to demonstrate the correctness, verifiability, privacy protection, and collusion resistance of PFLVA. Experimental results demonstrate that PFLVA maintains high model accuracy while significantly reducing the computational and communication overhead for participants compared to related schemes.

Byzantine-robust federated learning (FL) aims to obtain an accurate global model even with potentially Byzantine users. However, most existing schemes rely on measuring the overall differences between the entire gradient vectors of different users, which fail to effectively distinguish malicious gradients from benign ones caused by data heterogeneity under non-IID settings, thereby compromising model performance. To tackle this challenge, we propose BPFLH, a novel Byzantine-robust privacy preserving FL framework for heterogeneous data. BPFLH is the first to introduce Bray–Curtis dissimilarity into FL, capturing the element-wise differences among gradients from different users. This method reduces the risk of misclassifying benign gradi ents as malicious and enhance the model's robustness against Byzantine attacks in non-IID data environments. Furthermore, BPFLH leverages CKKS homomorphic encryption to protect local gradients, enabling secure aggregation and Byzantine user detection without compromising privacy. Extensive experiments on real-world datasets under various attack scenarios and data distributions demonstrate that BPFLH exhibits strong robustness against Byzantine attacks while preserving privacy and maintaining superior accuracy compared to existing Byzantine-robust FL methods, particularly in non-IID environments.

Unsupervised anomaly detection (UAD) methods typically detect anomalies by learning and reconstructing the normative distribution. However, since anomalies constantly invade and affect their surroundings, sub-healthy areas in the junction present structural deformations that could be easily misidentified as anomalies, posing difficulties for UAD methods that solely learn the normative distribution. The use of multimodal images can facilitate to address the above challenges, as they can provide complementary information of anomalies. Therefore, this paper propose a novel method for UAD in preoperative multimodal images, called Erasure Perception Diffusion model (EPDiff). First, the Local Erasure Progressive Training (LEPT) framework is designed to better rebuild sub-healthy structures around anomalies through the diffusion model with a two-phase process. Initially, healthy images are used to capture deviation features labeled as potential anomalies. Then, these anomalies are locally erased in multimodal images to progressively learn sub-healthy structures, obtaining a more detailed reconstruction around anomalies. Second, the Global Structural Perception (GSP) module is developed in the diffusion model to realize global structural representation and correlation within images and between modalities through interactions of high-level semantic information. In addition, a training-free module, named Multimodal Attention Fusion (MAF) module, is presented for weighted fusion of anomaly maps between different modalities and obtaining binary anomaly outputs. Experimental results show that EPDiff improves the AUPRC and mDice scores by 2% and 3.9% on BraTS2021, and by 5.2% and 4.5% on Shifts over the state-of-the-art methods, which proves the applicability of EPDiff in diverse anomaly diagnosis. The code is available at https://github.com/wjiazheng/EPDiff.

Graph neural networks (GNNs) are gaining popularity in diverse application domains and growing in complexity.As a result, it is crucial to achieve high-performance GNN execution.Among various techniques, algebraic transformations, including operator reordering and operator fusion, have been successfully applied to improve the computation and memory access efficiencies of DNN models.However,

Graph neural networks (GNNs) have proven to be powerful tools for learning from graph-structured data and have achieved great success in many applications. As the sizes of real-world graphs continue to grow, traditional GNN training methods face significant scalability challenges. Recently, disks have gained attention as a cost-effective solution to store large-scale graphs, and several disk-based GNN systems have been proposed to train large-scale graphs on a single machine. However, these systems either overlook the unique data characteristics of GNN workloads when designing cache plans or fail to fully exploit the multilevel hierarchy of storage and computation in system execution, thus resulting in disk I/O bottleneck and resource under-utilization. To address these issues, we present CaliEX, an advanced disk-based GNN system that employs joint optimizations of caching and execution within and across different training stages. CaliEX first designs tailored cache plans and execution policy for both graph topology and features to accelerate neighborhood sampling and feature gathering. Since these two training stages work on different types of data, CaliEX further auto-tunes the cache allocation and pipelines the execution across different stages to improve resource utilization and overall training throughput. Evaluations on multiple GNN models and various large-scale datasets show that CaliEX achieves 3.28 × speedup on average compared to existing disk-based GNN training systems.

Long context fine-tuning of large language models(LLMs) involves training on datasets that are predominantly composed of short sequences and a small proportion of longer sequences. However, existing approaches overlook this long-tail distribution and employ training strategies designed specifically for long sequences. Moreover, these approaches also fail to address the challenges posed by variable sequence lengths during distributed training, such as load imbalance in data parallelism and severe pipeline bubbles in pipeline parallelism. These issues lead to suboptimal training performance and poor GPU resource utilization. To tackle these problems, we propose a chunk-centric training method named ChunkFlow. ChunkFlow reorganizes input sequences into uniformly sized chunks by consolidating short sequences and splitting longer ones. This approach achieves optimal computational efficiency and balance among training inputs. Additionally, ChunkFlow incorporates a state-aware chunk scheduling mechanism to ensure that the peak memory usage during training is primarily determined by the chunk size rather than the maximum sequence length in the dataset. Integrating this scheduling mechanism with existing pipeline scheduling algorithms further enhances the performance of distributed training. Experimental results demonstrate that, compared with Megatron-LM, ChunkFlow can be up to 4.53x faster in the long context fine-tuning of LLMs. Furthermore, we believe that ChunkFlow serves as an effective solution for a broader range of scenarios, such as long context continual pre-training, where datasets contain variable-length sequences.

Long-context supervised fine-tuning (Long-SFT) plays a vital role in enhancing the performance of large language models (LLMs) on long-context tasks. To smoothly adapt LLMs to long-context scenarios, this process typically entails training on mixed datasets containing both long and short sequences. However, this heterogeneous sequence length distribution poses significant challenges for existing training systems, as they fail to simultaneously achieve high training efficiency for both long and short sequences, resulting in sub-optimal end-to-end system performance in Long-SFT. In this paper, we present a novel perspective on data scheduling to address the challenges posed by the heterogeneous data distributions in Long-SFT. We propose Skrull, a dynamic data scheduler specifically designed for efficient long-SFT. Through dynamic data scheduling, Skrull balances the computation requirements of long and short sequences, improving overall training efficiency. Furthermore, we formulate the scheduling process as a joint optimization problem and thoroughly analyze the trade-offs involved. Based on those analysis, Skrull employs a lightweight scheduling algorithm to achieve near-zero cost online scheduling in Long-SFT. Finally, we implement Skrull upon DeepSpeed, a state-of-the-art distributed training system for LLMs. Experimental results demonstrate that Skrull outperforms DeepSpeed by 3.76x on average (up to 7.54x) in real-world long-SFT scenarios.

Online GNN inference has been widely explored by applications such as online recommendation and financial fraud detection systems, where even minor delays can result in significant financial impact. Real-time dynamic graph sampling enables online GNN inference to reflect the latest graph updates in real-world graphs. However, online GNN inference typically demands millisecond-level latency Service Level Objectives (SLOs) as its performance guarantees, which poses great challenges for existing dynamic graph sampling approaches based on graph databases. The issues mainly arise from two aspects: long tail latency due to imbalanced data-dependent sampling and large communication overhead incurred by distributed sampling. To address these issues, we propose Helios, an efficient distributed dynamic graph sampling service to meet the stringent latency SLOs. The key ideas of Helios are 1) pre-sampling the dynamic graph in an event-driven approach, and 2) maintaining a query-aware sample cache to build the complete K-hop sampling results locally for inference requests. Experiments on multiple datasets show that Helios achieves up to 67× higher serving throughput and up to 32× lower P99 query latency compared to baselines.

Graph matching, as an important query technology, has been widely applied in various fields. With the increasing of graph data, users choose to encrypt a large number of graphs and store them in the cloud. Existing solutions to graph matching query over encrypted graphs require the user to execute a lot of time-consuming subgraph isomorphism (NP-complete problem) operations to extract the matched graphs, which inevitably brings heavy computational burden to the user. Therefore, how to reduce the number of subgraph isomorphisms is crucial for releasing the user from the heavy workload in a graph matching query scheme over encrypted graphs. In this paper, we propose a secure and efficient scheme for graph matching query over encrypted graphs. The main idea is to classify the query graph into frequent subgraph and infrequent subgraph, and adopt different strategies to perform the matching query. We design the novel secure index based on the frequent subgraphs and the edge labels to reduce the number of subgraph isomorphisms. When the query graph is a frequent subgraph, the proposed scheme can directly produce the exact result owing to this secure index. The user does not need to perform any subgraph isomorphism in this case. When the query graph is an infrequent subgraph, the proposed scheme can return a set of data graphs very close to the exact result. As a result, the proposed scheme reduces the number of subgraph isomorphisms substantially. Formal security proof is provided. Extensive experiments on real-world data sets show that the proposed scheme reduces nearly 90% subgraph isomorphism.

To verify the integrity of data stored in the cloud and improve storage efficiency, numerous cloud storage auditing schemes with deduplication have been proposed. In cloud storage, when users perform data dynamic operations, they should lose ownership of original data. However, existing schemes require re-encrypting the entire ciphertext when ownership changes and recalculating the authenticators for the blocks following the updated blocks when insertion or deletion operations are performed. These processes lead to high computation overhead. To address the above issues, we construct a secure deduplication and cloud storage auditing scheme with efficient dynamic ownership management and data dynamics. We adopt CAONT encryption method, where only a portion of the updated block is required to be re-encrypted during the ownership management phase, significantly reducing computation overhead. We also implement index switch sets to maintain the mapping between block indexes and cloud storage indexes of ciphertext blocks. By embedding cloud storage indexes within the authenticators, our scheme avoids the need to recalculate authenticators when users perform dynamic operations. Additionally, our scheme supports block-level deduplication, further improving efficiency. Through comprehensive security analysis and experiments, we validate the security and effectiveness of the proposed scheme.

Data integrity auditing with data deduplication allows the cloud to store only one copy of the identical file while ensuring the integrity of outsourced data. To facilitate flexible updates of outsourced data, data integrity auditing schemes supporting data dynamics and deduplication have been proposed. However, existing schemes either impose significant computation and communication burden to achieve data dynamics while ensuring data integrity and deduplication, or incur substantial computation overhead during the phases of authenticator generation and auditing. To address the above problems, in this paper, we construct a secure deduplication and efficient data integrity auditing scheme with data dynamics for cloud storage (DIADD). We design a lightweight authenticator structure to produce data authenticators for data integrity auditing, which can achieve authenticator deduplication and greatly reduce the computation overhead in the authenticator generation phase. Additionally, the time-consuming operations can be eliminated in the auditing phase. To enhance the efficiency of data dynamics, we employ the multi-set hash function technology to produce the file tags. This allows data owners to compute a new file tag without needing to recover the entire original file when performing dynamic operations. Furthermore, security analysis and experimental results demonstrate that DIADD is both secure and efficient.

Cloud storage auditing mechanism is used to check whether the data of users stored in the cloud is intact. Most existing auditing schemes for secure cloud storage are designed to check the integrity for specified files based on file identities. In some scenarios, the user would like to check the integrity of the files generated and uploaded to the cloud in a certain time period. Existing cloud storage auditing schemes cannot work well for supporting this practical requirement because the private information will be exposed. To satisfy this requirement, we propose a brand-new paradigm termed as privacy-preserving time-based auditing for secure cloud storage. The proposed paradigm allows the user to check whether the files generated and uploaded in a certain time period are intactly stored in the cloud. When intending to check the integrity of the files uploaded in this time period, the user only provides the challenged time period t to the Third Party Auditor (TPA). The TPA can verify the integrity of all the files based on this time period, but cannot know how many files and which files the user has generated and uploaded to the cloud in this time period. To decrease the complex overhead associated with certificate management, we introduce an identity-based auditing mechanism. We provide a specific security analysis to show the correctness, auditing soundness and privacy preserving of this scheme. The experiments demonstrate the efficiency of the proposed scheme.

Remote data integrity auditing, as a key technology for cloud storage, allows verifier to check cloud data integrity without downloading entire file from cloud server (CS). In practice, user might only concern the integrity of the files containing the specific keyword and would like to perform data dynamic operations on data. In this paper, we construct a practical keyword-based remote data integrity auditing scheme supporting full data dynamics. In such a scheme, a novel construction called keyword tag is designed. Using the keyword tag, TPA is able to simultaneously check whether CS correctly keeps all files containing the specific keyword. TPA can detect CS's misbehaviors once CS does not keep these files correctly. Furthermore, our scheme supports full data dynamics, including file-level updates and block-level updates. The keyword tag is updated when the user performs the file-level updates. To support block-level updates, we introduce an index switcher set to maintain the relationship between the block index and the authenticator index. We can avoid the recalculation of the authenticators by using authenticator indexes to generate authenticators. After cloud data are updated, data integrity still can be guaranteed. Security analysis and experimental results show that our scheme is provably secure and efficient.

As graph neural networks (GNNs) have achieved great success in many graph learning problems, it is of paramount importance to support their efficient execution. Different graphs and different operators present different patterns during execution. However, there is still a gap in the existing GNN acceleration research to explore adaptive parallelism. We show that existing GNN frameworks rely on handwritten static kernels, which fail to achieve the best performance across different graph operators and input graph structures. In this work, we propose uGrapher, a unified interface that achieves general high performance for different graph operators and datasets. The existing GNN frameworks can easily integrate our design for its simple and unified API. We take a principled approach that decouples a graph operator’s computation and schedule to achieve that. We first build a GNN-specific operator abstraction that incorporates the semantics of graph tensors and graph loops. We explore various schedule strategies based on the abstraction that can balance the well-established trade-off relationship between parallelism, locality, and efficiency. Our evaluation shows that uGrapher can bring up to 29.1× (3.5× on average) performance improvement over the state-of-the-art baselines on two studied NVIDIA GPUs.

Public auditing is a significant technique in cloud-based Internet of Things (IoT) systems, which enables the verifier to check the integrity of IoT data stored in the cloud. Nowadays, data become a core property for owners. Once the data of one owner are sold to another one, the ownership of these data has to be transferred. However, the existing public auditing schemes with data transfer require all the authenticators corresponding to the transferred data to be transformed to the new ones for integrity auditing. It incurs significant computation cost because of recomputing the new authenticators for all transferred data, especially when a vast quantity of data is being transferred. In addition, the data privacy and the identity privacy of the data owner cannot be protected for the verifier in such schemes. Thus, how to achieve efficient data transfer and privacy protection are key challenges in public auditing with data transfer for cloud-based IoT data. In this article, we propose a privacy-preserving identity-based public auditing scheme with efficient data transfer for cloud-based IoT data (PPADT). In PPADT, all the authenticators corresponding to the transferred data blocks do not need to be transformed. We only need to transform an aggregated authenticator in the integrity auditing phase. It means that the computation cost of data transfer is independent of the number of transferred data blocks. Furthermore, the data owner’s identity privacy can be ensured with the assistance of the private key generator. The data privacy can also be guaranteed by employing the random masking technique.

The key exposure is a serious threat for the security of data integrity auditing. Once the user's private key for auditing is exposed, most of the existing data integrity auditing schemes would inevitably become unable to work. To deal with this problem, we construct a novel and efficient identity-based data integrity auditing scheme with key-exposure resilience for cloud storage. This is achieved by designing a novel key update technique, which is fully compatible with BLS signature used in identity-based data integrity auditing. In our design, the Third Party Auditor (TPA) is responsible for generating update information. The user can update his private key based on the private key in one previous time period and the update information from the TPA. Furthermore, the proposed scheme supports real lazy update, which greatly improves the efficiency and the feasibility of key update. Meanwhile, the proposed scheme relies on identity-based cryptography, which makes certificate management easy. The security proof and the performance analysis demonstrate that the proposed scheme achieves desirable security and efficiency.

Surgical instrument segmentation is of great significance to robot-assisted surgery, but the noise caused by reflection, water mist, and motion blur during the surgery as well as the different forms of surgical instruments would greatly increase the difficulty of precise segmentation. A novel method called Branch Aggregation Attention network (BAANet) is proposed to address these challenges, which adopts a lightweight encoder and two designed modules, named Branch Balance Aggregation module (BBA) and Block Attention Fusion module (BAF), for efficient feature localization and denoising. By introducing the unique BBA module, features from multiple branches are balanced and optimized through a combination of addition and multiplication to complement strengths and effectively suppress noise. Furthermore, to fully integrate the contextual information and capture the region of interest, the BAF module is proposed in the decoder, which receives adjacent feature maps from the BBA module and localizes the surgical instruments from both global and local perspectives by utilizing a dual branch attention mechanism. According to the experimental results, the proposed method has the advantage of being lightweight while outperforming the second-best method by 4.03%, 1.53%, and 1.34% in mIoU scores on three challenging surgical instrument datasets, respectively, compared to the existing state-of-the-art methods. Code is available at https://github.com/SWT-1014/BAANet.

Graph neural networks (GNNs) have gained increasing popularity among researchers recently and have been employed in many applications. Training GNNs introduce a crucial stage called graph sampling. One of the most important sampling algorithms is Random Walk. However, Random Walk and many of its variants share and suffer from the same performance problem caused by random and fragmented memory access patterns, leading to significant system performance degradation. In this work, we present an efficient graph sampling engine on modern FPGAs integrated with in-package high bandwidth memory (HBM), which brings data closer and faster to the core logic. The hardware walker design is modular and easily scalable for massive parallelism, to fully utilize the available HBM channels. Our design also provides the flexibility to support Random Walk and two of its variants on both homogeneous and heterogeneous graphs. On real-world graph datasets, we achieve a 1.39 × -3.74 × speedup with a 2.42 × -6.69 × higher energy efficiency over highly optimized parallel baselines on a Xeon CPU. We also implement these algorithms on a NVIDIA Tesla VIOO GPU and achieve comparable dynamic energy consumption.

Using cloud storage services, users can store their data in the cloud to avoid the expenditure of local data storage and maintenance. To ensure the integrity of the data stored in the cloud, many data integrity auditing schemes have been proposed. In most, if not all, of the existing schemes, a user needs to employ his private key to generate the data authenticators for realizing the data integrity auditing. Thus, the user has to possess a hardware token (e.g., USB token, smart card) to store his private key and memorize a password to activate this private key. If this hardware token is lost or this password is forgotten, most of the current data integrity auditing schemes would be unable to work. In order to overcome this problem, we propose a new paradigm called data integrity auditing without private key storage and design such a scheme. In this scheme, we use biometric data (e.g., iris scan, fingerprint) as the user’s fuzzy private key to avoid using the hardware token. Meanwhile, the scheme can still effectively complete the data integrity auditing. We utilize a linear sketch with coding and error correction processes to confirm the identity of the user. In addition, we design a new signature scheme which not only supports blockless verifiability, but also is compatible with the linear sketch. The security proof and the performance analysis show that our proposed scheme achieves desirable security and efficiency.

Cloud storage auditing allows the users to store their data to the cloud with a guarantee that the data integrity can be efficiently checked. In order to release the user from the burden of generating data signatures, the proxy with a valid warrant is introduced to help the user process data in lightweight cloud storage auditing schemes. However, the proxy might be revoked or the proxy’s warrant might expire. These problems are common and essential in real-world applications, but they are not considered and solved in existing lightweight cloud storage auditing schemes. In this paper, we propose a lightweight identity-based cloud storage auditing scheme supporting proxy update, which not only reduces the user’s computation overhead but also makes the revoked proxy or the expired proxy unable to process data on behalf of the user any more. The signatures generated by the revoked proxy or the expired proxy can still be used to verify data integrity. Furthermore, our scheme also supports workload-based payment for the proxy. The security proof and the performance analysis indicate that our scheme is secure and efficient.

With cloud storage services, users can remotely store their data to the cloud and realize the data sharing with others. Remote data integrity auditing is proposed to guarantee the integrity of the data stored in the cloud. In some common cloud storage systems such as the electronic health records system, the cloud file might contain some sensitive information. The sensitive information should not be exposed to others when the cloud file is shared. Encrypting the whole shared file can realize the sensitive information hiding, but will make this shared file unable to be used by others. How to realize data sharing with sensitive information hiding in remote data integrity auditing still has not been explored up to now. In order to address this problem, we propose a remote data integrity auditing scheme that realizes data sharing with sensitive information hiding in this paper. In this scheme, a sanitizer is used to sanitize the data blocks corresponding to the sensitive information of the file and transforms these data blocks' signatures into valid ones for the sanitized file. These signatures are used to verify the integrity of the sanitized file in the phase of integrity auditing. As a result, our scheme makes the file stored in the cloud able to be shared and used by others on the condition that the sensitive information is hidden, while the remote data integrity auditing is still able to be efficiently executed. Meanwhile, the proposed scheme is based on identity-based cryptography, which simplifies the complicated certificate management. The security analysis and the performance evaluation show that the proposed scheme is secure and efficient.