Science China Information Sciences最新文献

The lower power object detection challenge (LPODC) at the IEEE/ACM Design Automation Conference is a premier contest in low-power object detection and algorithm (software)-hardware co-design for edge artificial intelligence, which has been a success in the past five years. LPODC focused on designing and implementing novel algorithms on the edge platform for object detection in images taken from unmanned aerial vehicles (UAVs), which attracted hundreds of teams from dozens of countries to participate. Our team SEUer has been participating in this competition for three consecutive years from 2020 to 2022 and obtained sixth place respectively in 2020 and 2021. Recently, we achieved the championship in 2022. In this paper, we presented the LPODC for UAV object detection from 2018 to 2022, including the dataset, hardware platform, and evaluation method. In addition, we also introduced and discussed the details of methods proposed by each year’s top three teams from 2018 to 2022 in terms of network, accuracy, quantization method, hardware performance, and total score. Additionally, we conducted an in-depth analysis of the selected entries and results, along with summarizing representative methodologies. This analysis serves as a valuable practical resource for researchers and engineers in deploying the UAV application on edge platforms and enhancing its feasibility and reliability. According to the analysis and discussion, it becomes evident that the adoption of a hardware-algorithm co-design approach is paramount in the context of tiny machine learning (TinyML). This approach surpasses the mere optimization of software and hardware as separate entities, proving to be essential for achieving optimal performance and efficiency in TinyML applications.

Accurate identification of malicious traffic is crucial for implementing effective defense counter-measures and has led to extensive research efforts. However, the continuously evolving techniques employed by adversaries have introduced the issues of concept drift, which significantly affects the performance of existing methods. To tackle this challenge, some researchers have focused on improving the separability of malicious traffic representation and designing drift detectors to reduce the number of false positives. Nevertheless, these methods often overlook the importance of enhancing the generalization and intraclass consistency in the representation. Additionally, the detectors are not sufficiently sensitive to the variations among different malicious traffic classes, which results in poor performance and limited robustness. In this paper, we propose intraclass consistency enhanced variational autoencoder with Class-Perception detector (ICE-CP) to identify malicious traffic under concept drift. It comprises two key modules during training: intraclass consistency enhanced (ICE) representation learning and Class-Perception (CP) detector construction. In the first module, we employ a variational autoencoder (VAE) in conjunction with Kullback-Leibler (KL)-divergence and cross-entropy loss to model the distribution of each input malicious traffic flow. This approach simultaneously enhances the generalization, interclass consistency, and intraclass differences in the learned representation. Consequently, we obtain a compact representation and a trained classifier for non-drifting malicious traffic. In the second module, we design the CP detector, which generates a centroid and threshold for each malicious traffic class separately based on the learned representation, depicting the boundaries between drifting and non-drifting malicious traffic. During testing, we utilize the trained classifier to predict malicious traffic classes for the testing samples. Then, we use the CP detector to detect the potential drifting samples using the centroid and threshold defined for each class. We evaluate ICE-CP and some advanced methods on various real-world malicious traffic datasets. The results show that our method outperforms others in identifying malicious traffic and detecting potential drifting samples, demonstrating outstanding robustness among different concept drift settings.

Covert communication has been widely investigated to avoid the transmission behavior being overheard by the warder. However, covert communication may be illegitimately utilized by unauthorized parties to evade the supervision of authorized agencies, which leads to great challenges to information security. To meet the need for authorized parties to monitor and prevent illegitimate transmission between unauthorized nodes, a novel paradigm, called legitimate monitor, is proposed for counter covert communications. In the preceding covert communication system, the covert transmission rate is the focus. Differently, the core concern of the legitimate monitor system is the outage probability of the transmission between unauthorized nodes, which should be maximized to interrupt the potential but undetectable transmission. To achieve these goals effectively, a proactive guarding approach is proposed, where the authorized warder detects the transmission behavior and emits jamming signals to interfere with the potential transmission, simultaneously. In particular, the jamming power at the warder is optimized under cases where the instantaneous/statistical channel state information is available. Besides, the corresponding outage probability is derived to evaluate the system performance, which can also be simplified to scenarios with a passive warder. Numerical results demonstrate that proactive guarding outperforms the passive one, especially when the warder is not proximal to the unauthorized transmitter. In addition, the proposed jamming power allocation scheme also outperforms other benchmark schemes.

Currently, a decision tree is the most commonly used data mining algorithm for classification tasks. While a significant number of studies have investigated privacy-preserving decision trees, the methods proposed in these studies often have shortcomings in terms of data privacy breach or efficiency. Additionally, these methods typically only apply to symmetric frameworks, which consist of two or more parties with equal privilege, and are not suitable for asymmetric scenarios where parties have unequal privilege. In this paper, we propose SecureCART, a three-party privacy-preserving decision tree training scheme with a privileged party. We adopt the existing pMPL framework and design novel secure interactive protocols for division, comparison, and asymmetric multiplication. Compared to similar schemes, our division protocol is 93.5–560.4 × faster, with the communication overhead reduced by over 90%; further, our multiplication protocol is approximately 1.5× faster, with the communication overhead reduced by around 20%. Our comparison protocol based on function secret sharing maintains good performance when adapted to pMPL. Based on the proposed secure protocols, we implement SecureCART in C++ and analyze its performance using three real-world datasets in both LAN and WAN environments. he experimental results indicate that SecureCART is significantly faster than similar schemes proposed in past studies, and that the loss of accuracy while using SecureCART remains within an acceptable range.

We first report a 2D material-based P-FET with excellent output current saturation characteristics and demonstrate the highest small-signal output impedance characteristics among all previously published 2D-FETs. Further, we utilize the excellent performance of the device to demonstrate a current mirror circuit, which has better high precision current replication performance than silicon-based devices. This work provides a possible technical approach for the development of high-performance analog circuit devices based on 2D materials.

Predicting future frames using historical spatiotemporal data sequences is challenging and critical, and it is receiving a lot of attention these days from academic and industrial scholars. Most spatiotemporal predictive algorithms ignore the valuable backward reasoning ability and the disparate learning complexities among different layers and hence, cannot build good long-term dependencies and spatial correlations, resulting in suboptimal solutions. To address the aforementioned issues, we propose a two-stage coarse-to-fine spatiotemporal predictive model with bidirectional distillation and level-specific meta-adaptation (SeeMore) in this paper, which includes a bidirectional distillation network (BDN) and a level-specific meta-adapter (LMA), to gain bidirectional multilevel reasoning. In the first stage, BDN concentrates on bidirectional dynamics modeling and coarsely constructs spatial correlations of different layers, while LMA is introduced in the second fine-tuning stage to refine the multilevel spatial correlations from a meta-learning perspective. In particular, BDN mimics the forward and backward reasoning abilities of humans in a distillation manner, which aids in the development of long-term dependencies. The LMA views learning of different layers as disparate but related tasks and guides the transfer of learning experiences among these tasks through learning complexities. Thus, each layer could be closer to its solutions and could extract more informative spatial correlations. By capturing the enhanced short-term spatial correlations and long-term temporal dependencies, the proposed model could extract adequate knowledge from sequential historical observations and accurately predict future frames whose backtracking preconditions are consistent with the historical sequence. Our work is general and robust enough to be integrated into most spatiotemporal predictive models without requiring additional computation or memory cost during inference. Extensive experiments on four widely used predictive learning benchmarks validated the proposed model’s effectiveness in comparison to state-of-the-art approaches (e.g., 10.6% improvement of Mean Squared Error on the Moving MNIST dataset).

This study proposes a model named BioKG-CMI to predict CMIs based on a biological knowledge graph. Faced with limited data, we employ subcellular localization to generate negative samples that align more closely with biological logic. To mine semantic information in circRNA and miRNA sequences, we introduce the pre-trained model BERT to learn sequence feature representation. Guided by the hypothesis that adjacent molecules have similar functions, we calculate spatial proximity between nodes of the same class. The DisMult algorithm is applied to extract the potential logical rules of the knowledge graph and learn entity and relationship representations. Subsequently, the integration of multi-feature successfully addresses the challenge of expressing the complex biological knowledge graph and overcoming the limitation of single-feature inadequacy. Multiple comparative experiments and case studies demonstrate the robustness of the proposed model.

Skill learning through reinforcement learning has significantly progressed in recent years. However, it often struggles to efficiently find optimal or near-optimal policies due to the inherent trial-and-error exploration in reinforcement learning. Although algorithms have been proposed to enhance skill learning efficacy, there is still much room for improvement in terms of skill learning performance and training stability. In this paper, we propose an algorithm called skill enhancement learning with knowledge distillation (SELKD), which integrates multiple actors and multiple critics for skill learning. SELKD employs knowledge distillation to establish a mutual learning mechanism among actors. To mitigate critic overestimation bias, we introduce a novel target value calculation method. We also perform theoretical analysis to ensure the convergence of SELKD. Finally, experiments are conducted on several continuous control tasks, illustrating the effectiveness of the proposed algorithm.

Team adaptation to new cooperative tasks is a hallmark of human intelligence, which has yet to be fully realized in learning agents. Previous studies on multi-agent transfer learning have accommodated teams of different sizes but heavily relied on the generalization ability of neural networks for adapting to unseen tasks. We posit that the relationship among tasks provides key information for policy adaptation. We utilize this relationship for efficient transfer by attempting to discover and exploit the knowledge among tasks from different teams, proposing to learn an effect-based task representation as a common latent space among tasks, and using it to build an alternatively fixed training scheme. Herein, we demonstrate that task representation can capture the relationship among teams and generalize to unseen tasks. Thus, the proposed method helps transfer the learned cooperation knowledge to new tasks after training on a few source tasks. Furthermore, the learned transferred policies help solve tasks that are difficult to learn from scratch.

Chiplet-based design, which breaks a system into multiple smaller dice (or “chiplets”) and reassembles them into a new system chip through advanced packaging, has received extensive attention in the post Moore’s law era due to its advantages in terms of cost, performance, and agility. However, significant challenges arise in this implementation approach, including the mapping of functional components onto chiplets, co-optimization of package and architecture, handling the increased latency of communication across functions in different dies, the uncertainty problems of fragment communication subsystems, such as maintaining deadlock-free when independently designed chiplets are combined. Despite various design approaches that attempt to address these challenges, surveying these approaches one-after-another is not the most helpful way to offer a comparative viewpoint. Accordingly, in this paper, we present a more comprehensive and systematic strategy to survey the various approaches. First, we divide them into chiplet-based system architecture design and interconnection design, and further classify them based on different architectures and building blocks of interconnection. Then, we analyze and cross-compare each classification separately, and in addition, we present a topical discussion on the evolution of memory architectures, design automation, and other relevant topics in chiplet-based designs. Finally, some discussions on important topics are presented, emphasizing future needs and challenges in this rapidly evolving field.