Large Language Models (LLMs) are being increasingly applied in various natural language processing tasks including safety-critical systems (e.g., medical diagnosis querying, and code generation for self-driving), where resilience to hardware transient faults is essential for guaranteed safety. Traditional Fault Injection (FI) approaches are time consuming due to a large number of repeated executions, which limits their scalability to fast evaluate large-scale resilience. To address these challenges, we propose LLM-IARE, a novel Input-Aware Resilience Estimation Model for LLMs under hardware transient faults. It takes advantage of twice static analysis and once dynamic execution to extract critical parameters to compute general resilience metrics such as Silent Data Corruption (SDC) rates. Fast static analysis can obtain the primary LLM parameters, while dynamic execution can provide input-sensitive attention profiling to characterize how input variations influence internal attention patterns dynamically. More importantly, our proposed LLM-IARE uses the obtained parameters for modeling at three levels (operation, module, and layer) so that the SDC rates of transient fault impacts on LLMs can be calculated quickly and accurately. Additionally, LLM-IARE is further extended to estimate the LLM application-level resilience metric, the cosine similarity reflecting the bit-upset induced semantic fault impacts on final output quality. Comprehensive experiments on six representative LLMs (for example, GPT-2, T5 and RoBERTa), and 30 BERT variants demonstrate that LLM-IARE achieves a fast and accurate LLMs resilience evaluation, with up to 7335× (average 4500×) speedup and an average logarithmic relative error of 3.94% compared with advanced LLVM-based fault injection methods. We further extend the evaluation to four larger Qwen2.5 models (0.5B–7B), where LLM-IARE maintains stable accuracy with logarithmic relative errors between 1.96% and 4.21% (average 3.39%).

As a significant global health threat, breast cancer needs accurate ultrasound images segmentation for early intelligent diagnosis. Medical image segmentation has made extensive use of the well-known CNNs, transformers, and their variations. However, because of the intrinsic restrictions resulting from the intense complexity of transformers and breast ultrasound pictures, high performance and fast speed cannot be accomplished simultaneously. The research proposes a unique BS-Mamba architecture for rapid and reliable tumor segmentation of breast ultrasound images. It incorporates three new modules to extract multi-level feature for improving the breast image segmentation performance and speed. On the one hand, the BreastSegMamba block effectively captures both global and local high-resolution information, and processes complex features fast. On the other hand, the hybrid attention block is designed to strengthen the extraction of positional and channel features for accurate segmentation. Additionally, the data-sensitive fuzzy logic function focuses on the edges of the ultrasound images to enhance features in critical areas for higher segmentation performance. Through comprehensive experiments and ablation studies on the BUSI and Dataset B breast ultrasound datasets, BS-Mamba significantly outperforms the current state-of-the-art methods in terms of segmentation accuracy. With a Dice coefficient of 92.13% and an IoU of 86.42% on the BUSI dataset and a Dice coefficient of 92.62% and an IoU of 86.67% on Dataset B, respectively, BS-Mamba demonstrated exceptional segmentation performance. Furthermore, in comparison to other established models, BS-Mamba achieves a training time acceleration of up to 2.88×, further substantiating its computational efficiency.

Hardware transient faults are proven to have a significant impact on deep neural networks (DNNs), whose safety-critical misclassification (SCM) in autonomous vehicles, healthcare, and space applications is increased up to four times. However, the inaccuracy evaluation using accurate fault injection is time-consuming and requires several hours and even a couple of days on a complete simulation platform. To accelerate the evaluation of hardware transient faults on DNNs, we design a unified and end-to-end automatic methodology, A-Mean, using the silent data corruption (SDC) rate of basic operations (such as convolution, addition, multiply, ReLU, and max-pooling) and a static two-level mean calculation mechanism to rapidly compute the overall SDC rate, for estimating the general classification metric accuracy and application-specific metric SCM. More importantly, a max-policy is used to determine the SDC boundary of non-sequential structures in DNNs. Then, the worst-case scheme is used to further calculate the enlarged SCM and halved accuracy under transient faults, via merging the static results of SDC with the original data from one-time dynamic fault-free execution. Furthermore, all of the steps mentioned above have been implemented automatically, so that this easy-to-use automatic tool can be employed for prompt evaluation of transient faults on diverse DNNs. Meanwhile, a novel metric “fault sensitivity” is defined to characterize the variation of transient fault-induced higher SCM and lower accuracy. The comparative results with a state-of-the-art fault injection method TensorFI+ on five DNN models and four datasets show that our proposed estimation method A-Mean achieves up to 922.80 times speedup, with just 4.20% SCM loss and 0.77% accuracy loss on average. The artifact of A-Mean is publicly available at https://github.com/breatrice321/A-Mean .

Due to the high-quality semantic information provided by the text modality, text-driven models have become the dominant approach for Multimodal Sentiment Analysis (MSA) in recent years. Despite notable progress in previous studies, two primary limitations remain: (i) aligning multimodal features often relies on simple matching of sequence length or feature dimension, which overlooks cross-modal heterogeneity. (ii) existing fusion techniques tend to over-rely on text, potentially diminishing the emotional data contributed by other modalities. To address these issues, in this paper, we propose a Dual-stage Normalization based Modality-Collaborative Fusion Network (DNMCN). Initially, to reduce modality discrepancies, we introduce a dual-stage normalization strategy, where features from different modalities were mapped into a common dimensional space in the first stage to facilitate effective cross-modal comparisons; sequence length inconsistencies caused by cropping and multiscale dimension reduction were addressed in the second stage. Additionally, to achieve high-quality cross-modal mapping without losing non-textual modality information, we propose an Adaptive modality-Collaborative Fusion Transformer (ACF-T) block. Specifically, in ACF-T block, textual semantics are first integrated into any non-text modality via multi-head attention. Next, a novel adaptive weighting strategy is introduced to balance the contribution of fused features and other non-textual modality features, thereby enhancing crossmodal interaction. Experimental results demonstrate that our method outperforms existing state-of-the-art approaches on the public benchmark datasets CH-SIMS, CMU-MOSI and CMU-MOSEI.

As an important online learning resource, Massive Open Online Courses have a large amount of comments, which can be exploited by aspect-level sentiment analysis to optimize MOOC teaching from different perspectives. However, there are two essential problems. One is that there is no open-source dataset on Chinese MOOC. The other problem is semantic information confusion caused by inherent polysemy of Chinese words and ambiguous expressions relatively relying on the context. To further characterize the special features of Chinese MOOC reviews, we build an open-source dataset with clean 5,000 MOOC reviews and propose a sentiment knowledge dependency tree–based graph neural network. The proposed model first uses the latest term frequency–inverse document frequency algorithm to extract high-frequency words and combines it with the Semantic Orientation Pointwise Mutual Information algorithm so a sentiment dictionary in the field of Chinese MOOCs is constructed. Then, the grammatical information of the dependency tree is merged with the sentiment knowledge information of the sentiment dictionary. Next, this novel model uses GCN to capture the long-distance feature information of the sentiment dependency tree and finally adopts the softmax function for sentiment classification. To further improve the model's performance, we also use BERT to enhance the text representation for higher accuracy. Meanwhile, the comparative experiments demonstrate that our proposed model takes advantages of the customized dependency tree by knowledge dictionary to achieve more accurate sentiment analysis than the state-of-the-art methods under different word embedding approaches.

Modern vessels integrate a massive digital infrastructure and navigation-dependent operating systems, allowing for ship-to-shore and ship-to-ship collaborative communication. However, the heightened interconnection of various maritime infrastructures inevitably amplifies the risk of vessel navigation and communication. Existing intrusion detection techniques were usually built on individual network events, failing to account for the multi-event long-term dependency problem caused by the high latency and low bandwidth of ship communication networks, therefore cannot tackle sophisticated cyber-ship attacks, resulting in lower accuracy in intrusion detection. In this paper, we propose a dependable Intrusion Detection System(IDS) based on Segment-wise CoTransformer(SegCoT) to detect cyber-ship intrusion events, which primarily contains a two-stage Network Pattern Extraction Component (NPEC) and an Intrusion Event Identification Component (IEIC). The NPEC automates the extraction of long-term dependency of massive intrusion events employing a SegEvent-wise Attention (SEA). Furthermore, the extracted dependencies are leveraged by the IEIC for specific intrusion type detection from a spatio-temporal feature fusion perspective. Based on a cyber-ship dataset collected from real ocean-going vessels, the proposed model achieves 99% intrusion detection accuracy, outperforming the existing state-of-the-art approaches.

Diverse breast masses in size, shape and place make accurate image segmentation more challenging in a unified deep-learning network.Therefore, based on the U-net network, an adaptive automatic whole-field segmentation network (AWDS-net) for characterising diverse breast masses is proposed to assist more accurate and fast medical diagnosis in this paper.In the encoder part of AWDS-net, a small mass extraction mechanism (SMEM) is designed to better retain fine-grained small mass location information, while a spatial pyramid module (SPM) is added to capture multi-scale context and high-resolution image information.In the decoder part, an attention gate (AG) mechanism is inserted to make the model automatically focus on the useful target region information, so that the extracted feature information can be used to build a symmetric encoderdecoder structure for automatic segmentation network of multiple masses in the full field of view.The experimental results on an opensource breast cancer dataset digital database for mammography (DDSM) show that compared with U-net, Attention-Unet, R2U-Net, and SegNet, the proposed AWDS-net achieves, up to higher image segmentation metrics of 3.16% accuracy, 20.59% sensitivity, 5.23% specificity,10.27%precision, 15.08% IoU and 14.21% F1-score with acceptable training time.

Due to the continuous technology scaling and lowering of operating voltages, modern computer systems are highly vulnerable to soft errors induced by the high-energy particles. Soft errors can corrupt program outputs leading to silent data corruption or a Crash. To protect computer systems against such failures, architects need to precisely and quickly identify vulnerable program instructions that need to be protected. Traditional techniques for program reliability estimation either use expensive and time-consuming fault injection or inaccurate analytical models to identify the program instructions that need to be protected against soft errors. In this work, we present GLAIVE, a graph learning-assisted model for fast, accurate, and transferable soft-error induced instruction vulnerability estimation. GLAIVE leverages a synergy between static analysis and data-driven statistical reasoning to automatically learn signatures of instruction-level vulnerabilities and their propagation to program outputs using a fine-grain error propagation information from the bit-level program graphs of a set of realistic benchmarks. Our experiments show that the learned knowledge of instruction vulnerability is transferable to unseen programs. We further show that GLAIVE can achieve an average 221× speedup and up to 33.09 % lower program vulnerability estimation error as compared to a baseline fault-injection technique, up to 30.29 % higher vulnerability estimation accuracy, and on average can cover up to 90.23 % vulnerable instructions for a given protection budget compared to a set of baseline machine learning algorithms.

With the exponential increase in the number of transistors under fast-paced technology progress, the soft error induced reliability issue is becoming even more challenging in heterogeneous multicore processor design. As there are significant opportunities to mitigate the soft error impacts through heterogeneous multicore scheduling, we show in this paper that the correlation among multiple applications exhibits important reliability characteristics, by defining a new metric to measure the system-level vulnerability factor of multiple applications and an approximate estimator to evaluate the metric fast and accurately for effective scheduling decisions. To approach these issues, we propose CASH, a Correlation-Aware Scheduling strategy to optimise heterogeneous multicore system reliability. Comprehensive simulation results demonstrate that the proposed approach is promising, achieving up to 21.4% reliability improvement with only 3.6% performance degradation when compared with performance-oriented scheduling policy.

Characterizing the soft error impacts is significant for a good trade off between design cost (e.g. Area and power) and reliability. In this paper, a heuristically mechanical model is proposed to quantify the soft error metric Architectural Vulnerability Factor (AVF) of storage structures (e.g. Register file and Cache) efficiently. This model not only considers the error spread among successive read operations, but also captures the logical masking effects of ALU unit, where errors propagate from the storage structure. The results of SPEC2000 INT demonstrate that, compared with the state-of-the-art method, the proposed model achieves a more accurate AVF estimation by up to 98.18% and on average 79.14% via only one more simulation for simple profiling.

Reliability issue, especially from transient errors due to scaling IC technology, low voltage supply, high frequency and heavy thermal effects, particles emission etc, has become a challenge for NoC design. Focus on this problem, an effective Reliability-Aware Arbitration Strategy simplified as RAPA, is proposed in this paper to decide which flits should be prioritized in the network transmission for higher application-level reliability. Different from pervious performance-oriented arbitration strategies, it includes the application-level reliability requirement to determine the reliability priority ranking. Flits patching mechanism is also used for avoiding starvation. The evaluation metric is redefined to emphasizing application-level reliability. Finally, we verify the reliability based prioritization policy on cycle accurate platform. And the simulation results show that the averaged successful delivery rate is upgraded from three nine of round robin (RR), old age based arbitration(OA) to five nine of our method RAPA. Especially, 67.15%, 41.83% reliability improvement in rest unreliable space on average are obtained over typical RR policy and OA based arbitration policy respectively with guaranteed performance.