Starting from the current state of generative artificial intelligence (AIGC) and large language models (LLM), I will first discuss the basic principles and shortcomings of the latest AIGC products, such as ChatGPT and Sora, along with their future improvements and development trends. I will mainly elaborate on the important roles and value of AIGC in the biopharmaceutical field. Recently, ChatGPT outperformed 17 doctors by accurately diagnosing arare disease in a 4-year-old boy. This demonstrates that, when applied appropriately, AI canindeed become an assistant in diagnosing and treating diseases. However, a study published in JAMA by Brigham and Women's Hospital, affiliated with Harvard University, showed that ChatGPT's cancer treatment recommendations were only completely accurate in 62‥ of cases, indicating that its results should be applied cautiously. One solution to this issue is the use ofcontent detection tools, such as AIGC-X and ZeroGPT. The vast information behind ChatGPT is an advantage, but in specialized fields, it also brings the downside of excessive interference information. To address this, our team has developed a large language model knowledge vector library system for autism that reduces training time and achieves similar objectives using only a small amount of training data. This lecture will also introduce the use of AIGC indesigning new drug molecules. By inputting numerous small drug molecules related to the treatment of a particular disease into the AIGC system, new drug molecules can be generated. Coupled with our powerful AI drug screening capabilities, we have the potential to design new drugs suitable for specific targets.
Edge-cloud networks connect a wide range of systems and devices, ranging from large data centers to small IoT devices. The potential advantages of adopting edge and cloud networking include the quick adaptation of new technologies, including a faster rollout and adoption of software and feature updates, as well as better management of various resources, including network and edge devices. This talk provides an overview of some challenges in efficient support for running AI/ML algorithms in edge-cloud networks. Our focus is on low latency, connectivity, and local data processing while still achieving efficiency. We will look attwo specific examples of AI/ML implementations: one is the optimal offloading of AI/ML code from IoT/edge devices to the cloud, and the other is exploring network topology and connectivity for efficient decentralized federated learning.
Starting from the current state of generative artificial intelligence (AIGC) and large language models (LLM), I will first discuss the basic principles and shortcomings of the latest AIGC products, such as ChatGPT and Sora, along with their future improvements and development trends. I will mainly elaborate on the important roles and value of AIGC in the biopharmaceutical field. Recently, ChatGPT outperformed 17 doctors by accurately diagnosing a rare disease in a 4-year-old boy. This demonstrates that, when applied appropriately, AI can indeed become an assistant in diagnosing and treating diseases. However, a study published in JAMA by Brigham and Women’s Hospital, affiliated with Harvard University, showed that ChatGPT's cancer treatment recommendations were only completely accurate in 62% of cases, indicating that its results should be applied cautiously. One solution to this issue is the use of content detection tools, such as AIGC-X and ZeroGPT. The vast information behind ChatGPT is an advantage, but in specialized fields, it also brings the downside of excessive interference information. To address this, our team has developed a large language model knowledge vector library system for autism that reduces training time and achieves similar objectives using only a small amount of training data. This lecture will also introduce the use of AIGC in designing new drug molecules. By inputting numerous small drug molecules related to the treatment of a particular disease into the AIGC system, new drug molecules can be generated. Coupled with our powerful AI drug screening capabilities, we have the potential to design new drugs suitable for specific targets.
Sarcasm is used on social media to humorously criticize or hurt specific individuals or things by expressing emotions opposite to the literal meaning of the text. Therefore, sarcasm detection is necessary to analyze people's true feelings. Recent research has proposed sarcasm detection methods using commonsense. Commonsense refers to the general information that humans possess and is knowledge helpful in understanding everyday situations. The existing sarcasm detection model, as focused on in this study, utilizes only one specific type of commonsense. However, verification using other types of commonsense has not been conducted. Additionally, although the existing model also proposes three explicit knowledge selection strategies based on sentiment polarity to narrow down the generated commonsense, the optimal strategies for different datasets are not identical. Therefore, it is necessary to obtain useful and appropriate commonsense information for sarcasm detection at the stage of commonsense generation. In this study, we compare the performance of sarcasm detection models using various types of commonsense and aim to select appropriate types of commonsense. We also make improvements in the input text to the commonsense generation model and post-processing after commonsense generation. Furthermore, we attempt to concatenate multiple types of commonsense for further performance improvement. The experimental results show that the model performance varies depending on the type of commonsense across three datasets. It is verified that acquiring useful commonsense for sarcasm detection through appropriate type selection can improve performance compared to the existing model.
WiFi backscatter can transmit its data by reflecting or not reflecting the WiFi signal. It can be well integrated into existing Internet of Things (IoT) devices to realize the transmission of tag data and the connection between WiFi devices. However, achieving high-rate backscatter communication on commercial WiFi devices is a challenge, especially without frequency shift. This paper proposes SymbolBack, a new backscatter communication system. By manually controlling the media access control (MAC) payload, we make data subcarriers of the transmitted orthogonal frequency division multiplexing (OFDM) symbols exactly the same. We set unique reference data for WiFi packets of different protocols, which only use a small number of symbols as reference symbols. We design a decoding algorithm based on the prediction of subsequent bits and the calculation of the Hamming distance. Tag bits are decoded by comparing the current symbol with all possible reference symbols. Decoding is done within the packet, thus the system only needs one receiver and no frequency shift. We validate our system on simulated environments and system prototypes. We show that our system achieves less than 0.1% bit error rate (BER) and a transfer rate of 60kbps at most.
Vertical Federated Learning (VFL) is an effective paradigm which matches the enterprises' demands of leveraging more valuable features to achieve better model performance while keeping their data decentralized. Training costs, especially communication costs among parties have always been one of the major bottlenecks in VFL. Almost all existing approaches is aimed at optimizing the overhead of a single VFL task. However, there are situations where a passive party cooperates with multiple active parties to jointly train VFL models task by task, which causing the passive party incurring significant duplication of training costs. In this scenario, because these methods tailored for single VFL task cannot leverage the inter-task correlations, they might be suboptimal in terms of optimization and the model's performance may also be less satisfactory. In this article, we focus on reducing the training overhead of the passive party in multiple VFL collaborations. We utilize all local data of the passive party, including data discarded in the VFL task, for unsupervised training to accelerate the convergence speed of VFL training. Some new problems arise due to the data heterogeneity between tasks and the issue of catastrophic forgetting in models. In order to solve these issues, we propose a training-efficient algorithm named MultiVFL for multiple VFL collaborations. MultiVFL adopt parameter-level knowledge retention and fine-tuning algorithm, motivated by the sparsity of the bottom model in VFL. Through this approach, MultiVFL is able to maintain efficient training across various scenarios. We conducted experiments on two datasets containing five tasks each, and the results indicate that MultiVFL can continuously and efficiently learn tasks in multiple rounds of VFL collaboration. Compared to traditional VFL methods, MultiVFL saves about 70% of training overhead on tabular datasets and 21% on image datasets with excellent model accuracy.
With the rapid advancement of space communication technology, the integrated space-air-ground-sea network demonstrates tremendous potential applications in various fields. Particularly in remote areas where traditional base station is almost unable to be established, satellite communication provides an optimal solution to address communication difficulties. Meanwhile, the continuous expansion of data scale makes the traditional end-to-end transmission mode of TCP/IP unsuitable to satellite communication. To address this issue, this paper proposes a multi-layer Satellite Internet of Things (SIoT) architecture based on Information-Centric Networking (ICN) and Software-Defined Networking (SDN) technologies, namely Information-Defined Satellite Network (IDSN). In the designs of satellite routing networks and ground-to-ground networks, the controller primarily integrates with ICN at the application layer, developing ICN-related services such as name resolution, name routing, and content caching to achieve intelligent data management. To enhance the compatibility of network infrastructure, IP protocol is reused, enabling effective differentiation between ICN request and IP request in the OpenFlow protocol. Layered and distributed controller deployment, and layered naming make the architecture more intelligent. Experimental results illustrate that the proposed framework can efficiently realize ICN caching features while simultaneously reducing latency.
The pervasive integration of deep neural networks (DNNs) within smart devices has significantly increased compu-tational workloads, consequently intensifying pressure on real-time performance and device power consumption. Offloading segments of DNNs to the edge has emerged as an effective strategy for reducing latency and device power usage. Nonethe-less, determining the workload to offload presents a complex challenge, particularly in the face of fluctuating device workloads and varying wireless signal strengths. This paper introduces a streamlined approach aimed at swiftly and accurately forecasting the computing latency of a DNN. Building upon this, an adaptive neurosurgeon framework is proposed to dynamically select the optimal partition point of a DNN during runtime, effectively minimizing computing latency. Through experimental validation, our proposed adaptive neurosurgeon demonstrates superior per-formance in reducing computing latency amidst changing DNN workloads across devices and varying wireless communication capabilities, outperforming existing state-of-the-art approaches, such as the autodidactic neurosurgeon.
Open-set domain adaptation transfers the model from a label-rich domain to a label-free domain that contains unknown-class samples. We incorporate active learning techniques to label a maximally informative subset of target samples to improve the model adaptation performance. Traditional active learning does not consider domain transfer, while active domain adaptation does not consider class mismatches. Therefore, these two methods cannot identify truly valuable samples in the case of open-set domain adaptation. To accommodate the two naturally different tasks of active learning and open-set domain adaptation, we explore those samples of known and unknown classes that are difficult to separate, called fuzzy samples. We propose two indicators to measure the domainness and fuzziness of samples. First, the source-target label alignment is used to calculate the domainness of the sample, and a candidate set is constructed for the non-low-domainness samples. Then, we use oversampling and clustering via the magnitude of feature update to select diverse subsets of fuzzy samples. Finally, open-set loss and information maximization loss are introduced to further promote the separation of known and unknown samples, and cross-entropy loss is introduced to classify known samples. Extensive experiments verify that our method remarkably exceeds the various types of baselines and achieves significant performance gains at a small cost of labeling. In some cases, our method can provide about 9.6% improvement over open-set domain adaptation methods and about 3% improvement over existing active learning methods.
Within cloud computing, Function as a Service (FaaS) deploys computations to serverless backends without exposing the underlying server environment. This facilitates high scalability and flexibility for software development. This innovative paradigm introduces a notable challenge: serverless service providers have limited resources, and the computing environment may not always be ready when a stateless function invocation request arrives. Thus, the users have to suffer from an initial latency known as cold start. To alleviate the cold start problem with resource constraints, it is important to manage the life cycle of functions' execution environment to ensure the resources consumption of living containers are under-limit. We present FISH, a Factor-Integrated Scheduler with IAT-Histogram, which enhances the HIST algorithm to improve efficient resource management in serverless computing environments. By employing a histogram of inter-arrival times (IAT) for optimizing pre-warm and keep-alive strategies, and incorporating an IAT-aware eviction policy that considers memory usage and invocation frequency, FISH aims to mitigate the cold start problem under resource constraints. Besides, we implement a simulator and evaluate FISH mainly with a fixed-keep-alive method with Azure Function Invocation Trace and obtain a maximum 32% increase in warm start rate and reduce total cold start time by up to 125% compared to the baseline.
Deep neural networks are vulnerable to adversar-ial examples, they can be deceived by adding imperceptible perturbations on benign inputs. Moreover, people exploit the transferability of adversarial examples to conduct black-box attacks, which means using adversarial examples constructed by surrogate models to deceive unknown models. In real-world scenarios, the network architecture and training dataset used by the victim model are unkonwn. The surrogate model and the victim model may employ different network structures and utilize distinct training datasets, thus impeding effective black-box attacks. While the current studies primarily focus on enhancing transferability across models, they have yet to achieve satisfactory transferability across both models and data domains simultaneously. In this paper, we employ a generative adversarial training framework which consists of a generator for constructing adversarial examples and a discriminator for identification. We innovatively introduce the concept of randomization throughout the entire training process to enhance the transferability of adversarial examples. Specifically, we employ random neural network weight pruning on the discriminator model during the training process, implicitly integrating a multitude of different discriminator models. This approach enhances the transferability of adversarial examples across models. Moreover, we apply random normalization to the dataset to address cross-domain transferability. Extensive experiments confirm the effectiveness of our method and adversarial examples generated using our approach exhibit high attack success rates in black-box attack scenarios.
While Machine Learning (ML) is widely used in many industries, one sector that is just beginning to leverage this technology is the legal field. Efforts like the construction of the OPP-115 Privacy Policy corpus make it possible to train Natural Language Processing (NLP) and ML models for legal use, and tools such as Polisis, Claudette, etc., which utilize these models can now provide users with descriptive annotations of various types of contracts, including Privacy Policies. However, these tools remain vulnerable to adversarial attacks, specifically adversarial legal documents, which differ from typical text based adversarial attacks due to the necessity of preserving the continuity of a legal document. Therefore, focusing on the OPP-115 corpus and Polisis specifically, we propose a framework that consists of a Doc2Vec model and a variety of shallow and deep learning methods for classification, followed by adversarial attacks on Privacy Policy segments which utilize GPT 3.5, a state-of-the-art LLM, in order to create an adversarial Privacy Policy. We evaluate the effectiveness of our method by comparing attack results to those produced by current methods, including Bert Attack and Text Fooler. Our procedure reveals not just that an attack using GPT 3.5 is an effective means of producing an adversarial policy, but that it is surprisingly easy to prompt a widely used state-of-the-art LLM to perform such a task.
Accurate bus arrival time prediction is of great practical significance for improving passenger travel experience and optimizing bus scheduling. Existing works address the problem mainly based on single bus line. However, these methods cannot capture the complex spatial-temporal dependencies of bus network, and have poor adaptability. To address this issue, we propose a DEep Bus Arrival Time Estimation network (DeBATE). Specifically, DeBATE first adopts two local representation learning components, station representation learning layer and road representation learning layer, to model the complex spatial-temporal dependencies within different bus lines between stations, as well as roads. Secondly, DeBATE utilizes a sequence learning component to incorporate line-specific global information into station and road representation. Finally, DeBATE adopts a multi-task learning component to simultaneously perform road segment prediction and global prediction to alleviate the accumulation of segment prediction errors. Experimental verification on 84 bus lines in Beijing demonstrates the effectiveness and superiority of DeBATE.
As activities related to cryptocurrency have been classified as illegal financial activities in China, many organizations have started to use encrypted connections to mine cryptocurrency. Traditional detection mechanism based on deep packet inspection is no longer effective for encrypted traffic. How to extract traffic fingerprint that can be used to detect encrypted mining traffic has become the main challenge of cryptocurrency mining detection. Although the TLS protocol encrypts data, the handshake packets of the TLS protocol contain rich plaintext information, which can be used to identify the encrypted traffic. It's common to use TLS related features to detect malicious encrypted traffic. But no study has been found to detect encrypted mining traffic based on TLS related features. Existing mining detection works mainly focus on Cryptojacking, and host-based method is commonly used. Only a few works adopt the detection mechanism based on network traffic. Most of these works mainly study the features of mining traffic in plaintext. The few remaining studies involving encryption simply mentioned the ability to detect encrypted traffic when talking about the robustness of their algorithm. This paper analyzes the features of encrypted mining traffic, particularly the features presented by TLS handshake. It was found that encrypted mining traffic exhibits clear differences in features such as certificate length, number of certificates, number of bytes sent from server to client, overall byte count, and length of certificate extensions when compared to normal traffic. 22 traffic features are used to construct fingerprint, which is then used to train the encrypted mining traffic classifier. Experiment shows that the encrypted mining traffic fingerprint proposed in this paper can detect encrypted mining traffic with high precision.
To deploy Transformer models on resource-limited edge devices, Hardware-Aware Neural Architecture Search (HWNAS) becomes a good choice due to the enormous computation of Transformers. In this paper, we propose a Precise and Transferable Latency Prediction method designed to accurately predict model latency on diverse platforms, thereby expediting the HWNAS search process. Initially, we train latency predictors on some platforms using collected data. Subsequently, for a new platform, we employ a latency exhibition pattern method to calculate similarity. By selecting the platform with the most comparable inference latency and transferring the trained latency predictor, we achieve higher prediction accuracy with less data. Experimental results demonstrate the effectiveness of our approach, boasting an average prediction accuracy of 97.2% across multiple platforms, all achieved with less than an hour of adaptation cost. Notably, our method outperforms state-of-theart approaches, showcasing an accuracy improvement of nearly 10% while reducing time costs by 20x
While machine learning has achieved tremendous success, the privacy security of its models faces challenges. To protect the privacy of machine learning models, researchers have proposed methods such as federated learning and differential privacy. However, the effectiveness of these methods in defending against attacks on model privacy at the practical level has not been comprehensively evaluated. In this paper, we focus on membership inference attacks targeting the privacy of machine learning models. By employing classical black-box membership inference attacks and a white-box membership inference attack proposed in this paper, we evaluate the privacy performance of the federated differential privacy framework in scenarios where privacy attacks are actively defended. Experimental results demonstrate that compared to centralized learning with differential privacy methods, models trained using the federated differential privacy framework exhibit stronger privacy performance and higher utility. We investigate the impact of differential privacy implementation mechanisms and privacy budgets on the privacy performance of federated learning models, providing insights and guidance for selecting critical privacy mechanisms and parameters in the practical application of the federated differential privacy framework.
Fall detection holds significant value in healthcare of elderly. However, the current fall detection algorithms based on millimeter-wave radar only detect whether an individual is fallen or not. This paper proposes a novel approach, utilizing the duration that an individual lies on the ground, to dynamically assess their physical posture. The fall detection problem is redefined as a human activity recognition task for further exploration. Nevertheless, current public scene-related datasets lack relevant activity data, such as falling and getting up. Therefore, in this paper we describe the approach of building a millimeter-wave radar point cloud dataset, encompassing nine activities for fall detection scene. Additionally, two methods for aligning data frame length are proposed during the data preprocessing stage. A human activity recognition model, Spatial Attention Temporal Attention Human Activity Recognition (SATA-HAR), is designed based on graph attention and transformer encoder. This model captures features in both spatial and temporal dimensions of activity data, and then utilizes a classifier for classification. The experimental results demonstrate that the methods and model proposed in this paper are highly effective. SATA-HAR achieves a recognition accuracy of 90.28%, surpassing other baseline models. It also shows strong training stability.
This paper proposes a phased application approach for intelligent identification and control methods, with a focus on the on-orbit risks and intelligent recognition and control technology of assembled spacecraft. In response to the significant risks and characteristics of assembled spacecraft, the paper proposes a small model-based relative risk identification method and a degradation reconstruction risk control method. These methods extract high-quality effective data from a small model perspective to conduct risk analysis of assembled spacecraft, identify key risks, and avoid the issue of disordered risk identification results. They also reduce system risks without increasing equipment redundancy or adding additional modules.
With the development of smart cities, particularly on campuses and in industrial parks, the rapid increase in the number of cameras has boosted the demand for multi-target, multi-camera tracking technology. However, current work on multi-target multi-camera tracking often involves a small number of cameras with limited monitoring coverage and short duration, insufficient for the analytical tasks required in smart parks. Moreover, existing works rely on high frame rate video streams as the data source, which, despite providing rich information, demand high storage and computational resources, limiting their application in long-term, large-scale trajectory tracking scenarios. In response to these challenges, we propose a new application scenario: large-scale multi-camera pedestrian tracking based on low-sampling-rate images. By utilizing low-sampling-rate surveillance images instead of videos and we have designed a Multi-dimensional Fusion Identity Verification architecture and a new trajectory tracking process to address the challenges brought about by large-scale camera tracking scenarios, such as environmental factors, diverse camera installation conditions, and the similarity in clothing among a large number of covered individuals. Our approach was tested in a real-world surveillance system with 2,456 cameras covering over 494.2 acres. Despite a significant reduction in data volume by over 60 times and an increase in camera count by 200 times compared to using video data sources of the real world such as WildTrack, our method still attained a pedestrian tracking accuracy of 75.8% and a recall rate of 57.4%. These results affirm our method's effectiveness and feasibility in large-scale environments.
In the complex space environment, targeting the characteristics of high integration and complexity in the new generation of spacecraft systems, and driven by the design needs of long lifespan and high reliability of spacecraft platforms, this paper proposes a system-level potential failure identification and analysis method based on the System-Theoretic Accident Model and Processes (STAMP). During the overall design process, this method initiates the subsystem reliability design requirements under space conditions. By constructing a system control logic architecture, the method analyzes potential unsafe control actions and determines the failure scenarios by inspecting each element in the control and feedback loops, proposing constraints on control processes to eliminate or mitigate the likelihood of failures, forming targeted subsystem reliability design requirements.
