With the rapid development of cloud computing, network architectures are moving towards distributed computing, which performs data processing at edge nodes to reduce latency, enabling more efficient and scalable network services. Nevertheless, this shift introduces significant security challenges due to the heterogeneity of communications protocols and the vulnerabilities of edge devices. To effectively secure these distributed networks, it is essential to perform multiple traffic analysis tasks, e.g. Network Intrusion Detection, Encrypted Traffic Classification, and Application Traffic Classification. However, existing methods have limited generic feature extraction and require the deployment of multiple models to solve multiple tasks, which exceeds the resource capacity of edge nodes. To address these challenges, we introduce a Lightweight Multitask Traffic Analysis Framework LiMTa, which novelly proposes a traffic pre-training method, FreqRec, and a lightweight multi-task model fine-tune method, MT-Adapter. FreqRec enables high-level semantic feature extraction by reconstructing the frequency features of traffic samples, and MT-Adapter efficiently performs multiple tasks by computing the pre-trained model only once. Experimental results demonstrate that our approach achieves state-of-the-art (SOTA) performance on six traffic analysis tasks. Moreover, the MT-Adapter module only fine-tunes a small number of parameters, accounting for only 6.37% of the pre-trained model’s parameters, and achieves the same result as the full fine-tuning. Compared to full fine-tuning, LiMTa reduces the time cost by 50.9% and the space cost by 57.4% in six edge traffic analysis tasks.

DDoS attacks have become one of the most severe cybersecurity threats, especially in application-layer attacks. With the emergence of Programmable Data Planes (PDPs), it has become possible to maintain line-rate throughput while achieving high detection rates, making them crucial in addressing DDoS challenges. However, due to the complexity of DDoS attacks, detection remains resource-intensive and overall network defense effectiveness is limited. This limitation becomes particularly pronounced in clustered environments, where coordinated defense is essential. This paper presents SISTAR, an innovative framework for efficient DDoS detection and mitigation using PDP. SISTAR integrates an improved Decision Tree - Constrained Threshold Segmentation (DT-CTS) model to achieve high detection accuracy while minimizing hardware resource usage. Through distributed deployment across multiple switches, SISTAR enhances network resilience by enabling rapid detection and coordinated response to DDoS attacks. We implement a prototype of SISTAR and evaluate its performance in a realistic testbed, the experimental results show that SISTAR surpasses existing models in terms of detection accuracy and resource efficiency. When combined with its alert pushback mechanism, SISTAR can effectively reduce network resource consumption caused by DDoS attacks.

Malicious traffic detection is a critical means to detect network attacks, and plays an essential role in ensuring network security. However, most existing methods only focus on single feature or statistical feature, failing to comprehensively learn traffic information. Moreover, feature extraction remains superficial, resulting in a lack of important details. Thus, this paper proposes a novel malicious traffic detection method, named “DeepBytes”, which consists of original traffic vectorization, feature extraction and fusion, and classification. The original traffic vectorization converts the traffic into vectors. In the feature extraction and fusion phase, we first use the proposed CB-ResNet network to extract deep byte feature, then effectively fuse flow, packet, and byte-level features through the proposed multi-attention algorithm. The traffic classification identifies whether the traffic is malicious and determines its specific type. Extensive experimental results on the ISCXIDS2012 and CICIDS2017 datasets indicate that the detection performance and false alarm of our proposed method achieve 99.96% and 0.01 %, outperforming the state-of-the-art methods. The results also demonstrate that our method has excellent generalization capability and can accurately detect unknown attacks.

The goal of the acoustic scene classification (ASC) task is to classify recordings into one of the predefined acoustic scene classes. However, in real-world scenarios, ASC systems often encounter challenges such as recording device mismatch, low-complexity constraints, and the limited availability of labeled data. To alleviate these issues, in this paper, a data-efficient and low-complexity ASC system is built with a new model architecture and better training strategies. Specifically, we firstly design a new low-complexity architecture named Rep-Mobile by integrating multi-convolution branches which can be reparameterized at inference. Compared to other models, it achieves better performance and less computational complexity. Then we apply the knowledge distillation strategy and provide a comparison of the data efficiency of the teacher model with different architectures. Finally, we propose a progressive pruning strategy, which involves pruning the model multiple times in small amounts, resulting in better performance compared to a single step pruning. Experiments are conducted on the TAU dataset. With Rep-Mobile and these training strategies, our proposed ASC system achieves the state-of-the-art (SOTA) results so far, while also winning the first place with a significant advantage over others in the DCASE2024 Challenge.

Abstract Spatio-temporal forecasting has wide applications across various domains, particularly in intelligent transportation systems, where it plays a crucial role. Traffic flow prediction, a typical spatio-temporal forecasting task, involves complex dependencies across both time and space dimensions. Current research predominantly relies on graph neural networks (GNNs) for modeling. However, deep GNN architectures often face the issue of over-smoothing. To address this challenge, recent studies have explored integrating residual connections or neural ordinary differential equations (ODEs) with GNNs. Nonetheless, existing graph ODE methods have limitations in initializing latent feature representations for time series data and capturing higher order spatio-temporal dependencies. Additionally, they struggle to extract multi-scale temporal dependencies. In this paper, we propose a framework called the Multiple Second-order Continuous Graph Neural Network. The framework utilizes a second-order continuous GNN, and experiments on four real-world datasets demonstrate that it outperforms mainstream baseline models, thereby confirming the effectiveness of the proposed method.

Machine anomalous sound detection is a useful technique for various applications, but it often suffers from poor generalization due to the challenges of data collection and complex acoustic environment. To address this issue, we propose a robust machine anomalous sound detection model that leverages self-supervised pre-trained models on large-scale speech data. Specifically, we assign different weights to the features from different layers of the pre-trained model and then use the working condition as the label for self-supervised classification fine-tuning. Moreover, we introduce a data augmentation method that simulates different operating states of the machine to enrich the dataset. Furthermore, we devise a transformer pooling method that fuses the features of different segments. Experiments on the DCASE2023 dataset show that our proposed method outperforms the commonly used reconstruction-based autoencoder and classification-based convolutional network by a large margin, demonstrating the effectiveness of large-scale pre-training for enhancing the generalization and robustness of machine anomalous sound detection. In Task2 of DCASE2023, we achieve 2nd place with these methods.

With the rapid development of 5G technology and deep learning, intelligent transportation systems (ITS) have been constantly improved and perfected. In the ITS, trajectory prediction, as a key part, requires the processing of a large amount of urban data, which places a high demand on computing resources. To reduce computational complexity and demand, this work proposes a lightweight spatial-temporal model based on convolutional neural networks which has fewer model parameters, faster prediction speed and higher prediction accuracy. This work transforms the complex trajectory into regular areas, reducing model computation while also more intuitively representing the change in trajectory. This work extracts multi-scale spatial features through spatial convolutions of different sizes to comprehensively understand the spatial variation of trajectory. Then, the temporal correlation of the trajectory is extracted through multi-layer time convolution. In addition, this paper introduces the trajectory relationship matrix(TRM) to enhance the global spatial relationship of the pedestrian trajectory and capture different spatial correlations for different areas. It analyzes the relationship between the areas where the trajectory is located by learning the historical trajectory. This work introduces extra factors such as pedestrian motion direction and speed to assist prediction, improving the model’s prediction ability and robustness. Experimental results show that the proposed MT-CNN model improves the accuracy by 22.7% compared to other models, with better performance and generalization ability, while also having excellent computational time and energy consumption.

HID attacks can imitate the keystrokes of human users to inject attack payloads to a victim computer. In this paper, we design a new HID attack tool that can inject a Trojan file into victim computer and especially simulate some user’s keystroke features. We also propose a new HID attack detection framework called HID Detector. It applies one-dimensional convolutional neural network to extract the abstract high-dimensional feature matrix of keystrokes, and uses the behavior feature differences to realize the detection of HID attacks. In addition, we have also deployed the detection model on a special hardware detection tool. The precision reaches 99.4% for detection task.

High-speed USB is prominently used in electronic devices, such as computers and mass storage devices, where sensitive data may be stored. There are risks of information leakage by conducted compromising emanations during data transmission with high-speed USB devices. Therefore, it is of great significance to research and analyze the security issues of high-speed USB for maintaining information security. However, little research is done on conducted compromising emanations of high-speed USB. In this paper, we propose a method to attack high-speed USB devices by conducted compromising emanations through USB hubs. First, we construct a USB hub-based attack model for high-speed USB devices. Then, we design a filtering mechanism with Kaiser filter to preprocess the collected leakage signal to obtain a cleaner leakage signal. Finally, by decoding and reconstructing the processed signal, the attack on the high-speed USB device is realized. Extensive experiments show that the method proposed in this paper can successfully attack high-speed USB devices, achieving a Bit-level Accuracy of 99.69% and a Character-level Accuracy of 98.04%. This demonstrates that our attack method is effective and verifies the existence of information leakage risk from high-speed USB devices through conducted compromising emanations in USB Hubs.

Traffic flow prediction is a vital pillar supporting the development and operation of modern transportation networks and is a significant research area within Intelligent Transportation Systems (ITS). The complexity of traffic scenarios and the heterogeneity of traffic data present greater challenges to traffic prediction research. Currently, most traffic flow prediction methods rely on Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), with Graph Convolutional Neural Networks (GCNs) being widely used. However, both homogeneous and heterogeneous GCNs have limitations in fully capturing the hierarchical and directional information of networks. To address this, we propose the Spatio-Temporal Heterogeneous Tree Convolutional Networks (STHTCN), a novel model for predicting future traffic flow. STHTCN utilizes a heterogeneously-structured tree to form spatial tree matrices, enabling the extraction of spatial characteristics and achieve the fusion of spatiotemporal features. Experimental results on two real-world traffic datasets from California demonstrate that the proposed STHTCN model reduces comprehensive error by an average of 14.43% and 10.38% compared to the heterogeneous graph convolutional baseline methods and exhibits excellent predictive capabilities.

The security of the Universal Serial Bus (USB) is critical due to its widespread use. Attackers can exploit electromagnetic emissions from USB devices to steal sensitive information, and detecting controlled emissions from USB has long been a challenging issue. Existing methods for detecting controlled emissions based on radiated compromising emanations are often hindered by environmental noise, electromagnetic interference, and device variability, leading to reduced detection capability. To address these challenges, this paper, for the first time, proposes a method for detecting controlled emissions from USB devices through conducted compromising emanations. First, we construct a dataset of USB leakage signals based on conducted compromising emanations. Then, we propose a multiscale time-frequency processing technique for USB leakage signals, applying time-frequency analysis to generate spectrograms containing both time-domain and frequency-domain features. By leveraging multiscale data augmentation, the accuracy and generalization of detection model are enhanced. Finally, we employ a Residual Network to automatically extract features and detect abnormal signals. Extensive experiments demonstrate that this method can effectively identify controlled emissions from USB devices, achieving an AUC of 99.52% and an ACC of 96.74%, providing a robust solution for anomaly detection and enhancing the security of USB devices.

The complexity of traffic scenarios, the spatial-temporal feature correlations pose higher challenges for traffic prediction research. Traffic spatial-temporal model is an essential method in this research field, primarily focusing on capturing the spatial-temporal features among nodes and their neighboring nodes. However, existing methods lack comprehensive consideration of directional and hierarchical features among traffic nodes. They are mostly applicable to scenarios with random uniform distribution of nodes, but not suitable for more complex small-scale aggregation distribution scenarios. Therefore, this study proposes the Tree Convolutional Network (TreeCN), a tree-based structure. The data design and model design of TreeCN focus on capturing the directional and hierarchical features among nodes. The directional and hierarchical relationships among nodes are represented by the plane tree matrix and constructed as the spatial tree matrix. The TreeCN, with a full convolution network, performs a bottom-up convolution structure on the tree matrix to complete the task of node feature capturing. In this study, TreeCN is thoroughly compared with statistical, machine learning, and deep learning methods in traffic time series prediction. The experimental results show that TreeCN not only performs well in scenarios with random uniform distribution but also exhibits outstanding effect in more complex small-scale aggregation distribution. Moreover, TreeCN adheres to the design principles of Graph Convolutional Networks (GCN) in capturing the spatial features of traffic nodes and can further capture directional and hierarchical features among them. This is expected to make TreeCN a new method to handle complex traffic scenarios and improve prediction accuracy.

Traffic flow forecasting plays a crucial role in Intelligent Transportation Systems (ITS) for the development and operation of modern transportation networks. Current methods primarily rely on Graph Convolutional Neural Networks (GNN) and Recurrent Neural Networks (RNN) to predict traffic flow. However, these methods face challenges in effectively capturing hierarchical and directional information within the traffic network while quantitatively balancing the relationships between current, previous, and future time data. To address these issues, this paper introduces a novel approach called Spatio-Temporal Tree and Gauss Convolutional Network (ST-TGCN) for traffic flow forecasting. The model utilizes a tree structure to construct a planar tree matrix for extracting spatial features and employs gaussian temporal convolution to extract temporal features of traffic flow. Experimental results demonstrate that ST-TGCN outperforms baseline methods, indicating its superior predictive capabilities.

Abstract Taxi flow forecast is significant for planning transportation and allocating basic transportation resources. The flow forecast in the urban adjacent area is different from the fixed-point flow forecast. Their data are more complex and diverse, which make them more challenging to forecast. This paper introduces a deep spatial–temporal forecast (DeepSTF) model for the flow forecasting of urban adjacent area, which divides the urban into grids and makes it have a graph structure. The model builds a spatial–temporal calculation block, which uses graph convolutional network to extract spatial correlation feature and uses two-layer temporal convolutional networks to extract time-dependent feature. Based on the theory of dilation convolution and causal convolution, the model overcomes the under-fitting phenomenon of other models when calculating with rapidly changing data. In order to improve the accuracy of prediction, we take weather as an implicit factor and let it participate in the feature calculation process. A comparison experiment is set between our model and the seven existing traffic flow forecast models. The experimental results prove that the model has better the capabilities of long-term traffic prediction and performs well in various evaluation indicators.

Traffic flow prediction has always been the focus of research in the field of Intelligent Transportation Systems, which is conducive to the more reasonable allocation of basic transportation resources and formulation of transportation policies. The spread of COVID-19 has seriously affected the normal order in the transportation sector. With the increase in the number of infected people and the government's anti-epidemic policy, human outgoing activities have gradually decreased, resulting in increasingly obvious discreteness and irregularities in traffic flow data. This article proposes a deep-space time traffic flow prediction model based on discrete wavelet transform (DSTM-DWT) to overcome the highly discrete and irregular nature of the new crown epidemic. First, DSTM-DWT decomposes traffic flow into discrete attributes, such as flow trend, discrete amplitude, and discrete baseline. Second, we design the spatial relationship of the transportation network as a graph and integrate the new crown pneumonia epidemic data into the characteristics of each transportation node. Then, we use the graph convolutional network to calculate the spatial correlation of each node, and the temporal convolutional network to calculate the temporal correlation of the data. In order to solve the problem of high discreteness of traffic flow data during the epidemic, this article proposes a graph memory network (GMN), which is used to convert discrete magnitudes separated by discrete wavelet transform into high-dimensional discrete features. Finally, use DWT to segment the predicted traffic data, and then perform the inverse discrete wavelet transform between the newly segmented traffic trend and discrete baseline and the discrete model predicted by GMN to obtain the final traffic flow prediction result. In simulation experiments, this work was compared with the existing advanced baselines to verify the superiority of DSTM-DWT.

Mobile edge computing (MEC) has become a research trend that solves effectively computationally intensive and latency-sensitive tasks. MEC environments in the real world are dynamic and uncertain and then the changes of the environments bring challenges to the generalization and robustness of offloading algorithms. In order to solve the above problem, we propose a meta-reinforcement learning task offloading algorithm GASTO based on Graph Neural Network and seq2seq network. Meta-learning can learn the optimal initialization parameter through several gradient descent steps and samples to adapt to new environments more quickly. The task generated in the user equipment is composed of multiple subtasks rather than a single task, and there are dependencies between the subtasks. Therefore, the task on the user equipment is modeled as a Directed Acyclic Graph (DAG). The connection relationship between the subtasks in DAG plays an important role. Drawing on the idea of message passing, Graph Neural Network is applied in DAG to extract the intrinsic correlation between subtasks in GASTO. In addition, Seq2Seq network can reduce the dimension of action space effectively, and the scheduling decisions of all subtasks can be generated simultaneously. Besides, in order to enhance the sampling efficiency of tasks and the robustness of GASTO, the priority of sampling tasks is adjusted dynamically during the training process. The experimental results of four algorithms in different environments show that the proposed algorithm GASTO can quickly adapt to the new environment.

This work presents our systems for the ADD2022 challenge. The ADD2022 challenge is the first audio deep synthesis detection challenge, which aims to spot various kinds of fake audios. We have explored using unsupervised pretraining models to build fake audio detection systems. Results indicate that unsupervised pretraining models can achieve excellent performance for fake audio detection. Our final EER results for low-quality fake audio detection and partially fake audio detection are 32.80% and 4.80% relatively. For partially fake audio detection, our results ranked first in the competition. Even trained with totally mismatched data, our method still generalizes well for partially fake audio detection.

In this paper, we present Multi-scale Feature Aggregation Conformer (MFA-Conformer), an easy-to-implement, simple but effective backbone for automatic speaker verification based on the Convolution-augmented Transformer (Conformer).The architecture of the MFA-Conformer is inspired by recent stateof-the-art models in speech recognition and speaker verification.Firstly, we introduce a convolution subsampling layer to decrease the computational cost of the model.Secondly, we adopt Conformer blocks which combine Transformers and convolution neural networks (CNNs) to capture global and local features effectively.Finally, the output feature maps from all Conformer blocks are concatenated to aggregate multi-scale representations before final pooling.We evaluate the MFA-Conformer on the widely used benchmarks.The best system obtains 0.64%, 1.29% and 1.63% EER on VoxCeleb1-O, SITW.Dev, and SITW.Eval set, respectively.MFA-Conformer significantly outperforms the popular ECAPA-TDNN systems in both recognition performance and inference speed.Last but not the least, the ablation studies clearly demonstrate that the combination of global and local feature learning can lead to robust and accurate speaker embedding extraction.We have also released the code 1 for future comparison.

By the prediction of traffic conditions, the occurrence of traffic congestion can be warned in advance, so that the traffic managers can intervene in time, which can help to reduce the risk of traffic congestion. Therefore, aiming at the problem of traffic congestion, a prediction method for dual traffic condition indicators is proposed. The method for capturing spatial dependence based on the topology of roads and road driving direction is proposed to provide more flexible and targeted spatial features for predicting traffic conditions. In addition, according to the real-time and accuracy requirements of traffic conditions prediction, a novel model named dual-channel convolution block is designed to capture the temporal dependence of traffic conditions. Learning from the idea of ensemble learning,$K$independent base models are trained to predict traffic condition at the same time, and a model fusion mechanism based on real-time traffic conditions is proposed to fuse the predictions of the base models so that the model can have stronger generalization ability to adapt to various noise data in real traffic conditions. The proposed method is validated on the traffic data sets and compares with the optimal model of all the existing models, the proposed method reduces MAPE of speed prediction by 12.1% and TTI prediction by 10.4%.

With the continuous growth of urban population, it is urgent for people to accurately plan the travel time. Therefore, travel time prediction of urban areas has become a key research direction in the field of smart cities. At present, several studies on travel time prediction are only conducted on a single mode, where the prediction process only treats a certain vehicle as an isolated traffic state on the route. However, the factors affecting traffic are extremely complex, thus making it very difficult to produce a comprehensive forecast. Based on this situation, the mixed existing model and mutual influence of multiple modes of transportation in the city are fully considered, and a multimodal deep learning model namely MC-GRU (Multimodal Convoluted Gated Recurrent Unit Network) is proposed. At the same time, to solve the problem of some objective factors, such as departure time and travel distance, we propose an attribute module to deal with these implicit factors. In addition, to explore the interaction between different modes of vehicles, a feature fusion module for obtaining the interaction effect between different modes of vehicles is proposed. Finally, we use GRU to learn the long-term dependence. MC-GRU can realize the accurate prediction of travel time in multimodal traffic state, as well as implement travel time prediction for three types of travel modes. The experimental results show that MC-GRU achieves higher prediction accuracy on a challenging real world dataset as compared with MAE, MAPE and RMSE.

Pervasive needs for data explorations at all scales have populated modern distributed platforms with workloads of different characteristics. The growing complexities and diversities have thereafter imposed distinct challenges to execute them on shared clusters in corporate or public clouds. This paper presents Fangorn, an adaptive execution framework built on an enriched graph model. As the underlying infrastructure for core computation platforms at Alibaba, Fangorn supports various execution modes and caters to heterogeneous workloads. With the capability to orchestrate graph executions with both long-running and requested-on-demand resources at the same time, Fangorn allows exploration of tradeoffs between latency and resource efficiency, for jobs of all scales. By modeling distributed job executions as mutable graphs with pluggable components, Fangorn offers a systematic framework to adjust job executions adaptively, according to data statistics collected during run-time. Fangorn supports an array of different computation engines ranging from relational to deep learning, and is fully deployed on production clusters across Alibaba. It manages tens of millions of distributed jobs daily, with job size scaling from one to half-million.

Blind people have intelligent tools to rely on for travel with the development of navigation technology. The GPS navigation, blind track, etc., are tools that blind people often use when traveling outdoors. However, indoor navigation tools and technology for blind people are lacking. We propose an obstacle avoidance algorithm and a spatial‐temporal model of trajectory prediction for the indoor travel task of the blind. The focus of this work is that it enables the blind to accurately avoid obstacles and achieve high accuracy trajectory prediction aiming at the unique movement characteristics of the blind. We set up a variety of baselines to conduct an experimental evaluation on a dataset of blind trajectories in a multistorey shopping mall. The experimental results show the advantages of the data model and predictive model of this work.

Hardware design, especially the very large scale integration(VLSI) and systems on chip design(SOC), utilizes many codes from third-party intellectual property (IP) providers and former designers. Hardware Trojans (HTs) are easily inserted in this process. Recently researchers have proposed many HTs detection techniques targeting the design codes. State-of-art detections are based on the testability including Controllability and Observability, which are effective to all HTs from TrustHub, and advanced HTs like DeTrust. Meanwhile, testability based detections have advantages in the timing complexity and can be easily integrated into recently industrial verification. Undoubtedly, the adversaries will upgrade their designs accordingly to evade these detection techniques. Designing a variety of complex trojans is a significant way to perfect the existing detection, therefore, we present a novel design of HTs to defeat the testability based detection methods, namely DeTest. Our approach is simple and straight forward, yet it proves to be effective at adding some logic. Without changing HTs malicious function, DeTest decreases controllability and observability values to about 10% of the original, which invalidates distinguishers like clustering and support vector machines (SVM). As shown in our practical attack results, adversaries can easily use DeTest to upgrade their HTs to evade testability based detections. Combined with advanced HTs design techniques like DeTrust, DeTest can evade previous detecions, like UCI, VeriTrust and FANCI. We further discuss how to extend existing solutions to reduce the threat posed by DeTest.

In this paper, in order to improve keyword spotting (KWS) performance in a live broadcast scenario, we propose to use a template matching method based on acoustic word embeddings (AWE) as the second stage to verify the detection from the Deep KWS system. AWEs are obtained via a deep bidirectional long short-term memory (BLSTM) network trained using limited positive and negative keyword candidates, which aims to encode variable-length keyword candidates into fixed-dimensional vectors with reasonable discriminative ability. Learning AWEs takes a combination of three specifically-designed losses: the triplet and reversed triplet losses try to keep same keyword candidates closer and different keyword candidates farther, while the hinge loss is to set a fixed threshold to distinguish all positive and negative keyword candidates. During keyword verification, calibration scores are used to reduce the bias between different templates for different keyword candidates. Experiments show that adding AWE-based keyword verification to Deep KWS achieves 5.6% relative accuracy improvement; the hinge loss brings additional 5.5% relative gain and the final accuracy climbs to 0.775 by using calibration scores.

Proxy-word based out of vocabulary (OOV) keyword search has been proven to be quite effective in keyword search. In proxy-word based OOV keyword search, each OOV keyword is assigned several proxies and detections of the proxies are regarded as detections of the OOV keywords. However, the confidence scores of these detections are still those of the proxies from lattices. To obtain a better confidence measure, we employ an LSTM-CTC verification method in this work and the confidence scores are regenerated. OOV keyword search results on the evalpart1 dataset of the OpenKWS16 Evaluation have shown consistent improvement and the maximum relative improvement can reach 21.06% for the MWTW metric.

This paper presents the Swahili keyword search system developed by the THUEE team for the OpenKWS15 evaluation, which is conducted by NIST under the IARPA Babel program. There are several highlights in the development of the system, including automatic generation of the pronunciation lexicon, aggressive data augmentation, the multilingual bottleneck feature extractor trained from 6 languages, text selection from web data for language model training, semi-supervised training for acoustic models and language models, out-of-vocabulary keyword detection using morphemes and a rich diversity of the systems for combination. A wide variety of acoustic modeling techniques are explored and compared. Up to 12 different individual systems are used for combination. The system achieves the state-of-the-art performance in the required condition of the evaluation.

It has been demonstrated that system fusion can significantly improve the performance of keyword search. In this paper, we compare the performance of several widely-used arithmetic-based fusion methods using different normalization pipeline and try to find the best pipeline. A novel arithmetic-based fusion method is proposed in this work. The method supplies a more effective way to incorporate the number of systems which have non-zero scores for a detection. When tested on the development test dataset of the OpenKWS15 Evaluation, the proposed method achieves the highest maximum term-weighted value (MTWV) and actual term-weighted value (ATWV) among all other arithmetic-based fusion methods. Usually, discriminative fusion methods employing classifiers can outperform arithmetic-based fusion methods. A DNN-based fusion method is explored in this work. After word-burst information is added, the DNN-based fusion method outperforms all other methods. In addition, it is notable that our arithmetic-based method achieves the same MTWV as the DNN-based method.

The OpenKWS14 keyword search evaluation is one of the most challenging and influential evaluations in the field of speech recognition. Its goal is to build a high-performance keyword search system for a minority language with limited training data in a short period of time. We present the system of the Department of Electronic Engineering, Tsinghua University (THUEE team) for the OpenKWS14 keyword search evaluation. The highlights of the system include the use of convolutional maxout neural networks for acoustic modeling and the use of neural network language models for one-pass lattice generation. The final system is a fusion of 8 sub-systems. The system has achieved an actual term weighted value (ATWV) of 0.5107 for the full language pack (FullLP) condition in the evaluation, ranking third among the participating teams.

Semi-supervised graph-based band selection methods have shown satisfying performances to choose the valuable bands for the hyper spectral data classification in case of very limited labeled samples. However, the calculation of adjacency matrices based on all labeled and unlabeled samples requires a large computational load which can be unacceptable with the huge amounts of unlabeled samples available. To address the problem, an unlabeled sample reduction method is proposed. The method involves dimensional reduction through PCA, over-segmentation through watershed, random sample selection from the resulting clusters. The band selection and classification experiments on hyper spectral data demonstrate that the proposed method can help improve the computational efficiency and performances of the graph-based algorithms by choosing the representative samples.