Biblio

To exploit high temporal correlations in video frames of the same scene, the current frame is predicted from the already-encoded reference frames using block-based motion estimation and compensation techniques. While this approach can efficiently exploit the translation motion of the moving objects, it is susceptible to other types of affine motion and object occlusion/deocclusion. Recently, deep learning has been used to model the high-level structure of human pose in specific actions from short videos and then generate virtual frames in future time by predicting the pose using a generative adversarial network (GAN). Therefore, modelling the high-level structure of human pose is able to exploit semantic correlation by predicting human actions and determining its trajectory. Video surveillance applications will benefit as stored “big” surveillance data can be compressed by estimating human pose trajectories and generating future frames through semantic correlation. This paper explores a new way of video coding by modelling human pose from the already-encoded frames and using the generated frame at the current time as an additional forward-referencing frame. It is expected that the proposed approach can overcome the limitations of the traditional backward-referencing frames by predicting the blocks containing the moving objects with lower residuals. Our experimental results show that the proposed approach can achieve on average up to 2.83 dB PSNR gain and 25.93% bitrate savings for high motion video sequences compared to standard video coding.

Advanced video compression is required due to the rise of online video content. A strong compression method can help convey video data effectively over a constrained bandwidth. We observed how more internet usage for video conferences, online gaming, and education led to decreased video quality from Netflix, YouTube, and other streaming services in Europe and other regions, particularly during the COVID-19 epidemic. They are represented in standard video compression algorithms as a succession of reference frames after residual frames, and these approaches are limited in their application. Deep learning's introduction and current advancements have the potential to overcome such problems. This study provides a deep learning-based video compression model that meets or exceeds current H.264 standards.

With the rapid development of multimedia and short video, there is a growing concern for video copyright protection. Some work has been proposed to add some copyright or fingerprint information to the video to trace the source of the video when it is stolen and protect video copyright. This paper proposes a video watermarking method based on a deep neural network and curriculum learning for watermarking of sliced videos. The first frame of the segmented video is perturbed by an encoder network, which is invisible and can be distinguished by the decoder network. Our model is trained and tested on an online educational video dataset consisting of 2000 different video clips. Experimental results show that our method can successfully discriminate most watermarked and non-watermarked videos with low visual disturbance, which can be achieved even under a relatively high video compression rate(H.264 video compress with CRF 32).

In this work we propose a novel deep learning approach for ultra-low bitrate video compression for video conferencing applications. To address the shortcomings of current video compression paradigms when the available bandwidth is extremely limited, we adopt a model-based approach that employs deep neural networks to encode motion information as keypoint displacement and reconstruct the video signal at the decoder side. The overall system is trained in an end-to-end fashion minimizing a reconstruction error on the encoder output. Objective and subjective quality evaluation experiments demonstrate that the proposed approach provides an average bitrate reduction for the same visual quality of more than 60% compared to HEVC.

The requirements of much larger file sizes, different storage formats, and immersive viewing conditions pose significant challenges to the goals of compressing VR content. At the same time, the great potential of deep learning to advance progress on the video compression problem has driven a significant research effort. Because of the high bandwidth requirements of VR, there has also been significant interest in the use of space-variant, foveated compression protocols. We have integrated these techniques to create an end-to-end deep learning video compression framework. A feature of our new compression model is that it dispenses with the need for expensive search-based motion prediction computations by using displaced frame differences. We also implement foveation in our learning based approach, by introducing a Foveation Generator Unit (FGU) that generates foveation masks which direct the allocation of bits, significantly increasing compression efficiency while making it possible to retain an impression of little to no additional visual loss given an appropriate viewing geometry. Our experiment results reveal that our new compression model, which we call the Foveated MOtionless VIdeo Codec (Foveated MOVI-Codec), is able to efficiently compress videos without computing motion, while outperforming foveated version of both H.264 and H.265 on the widely used UVG dataset and on the HEVC Standard Class B Test Sequences.

Real-time data transmissions from a vehicle enhance road safety and traffic efficiency by aggregating data in a central server for data analytics. When drivers share their instantaneous vehicular information for a service provider to perform a legitimate task, a curious service provider may also infer private information it has not been authorized for. In this paper, we propose a privacy preservation framework based on the Hilbert Schmidt Independence Criterion (HSIC) to sanitize driving data to protect the vehicle’s trajectory from adversarial inference while ensuring the data is still useful for driver behavior detection. We develop a deep learning model to learn the HSIC sanitizer and demonstrate through two datasets that our approach achieves better utility-privacy trade-offs when compared to three other benchmarks.

This paper studies Distributed Denial of Service (DDoS) attack detection by adopting the Deep Neural Network (DNN) model in Software Defined Networking (SDN). We first deploy the flow collector module to collect the flow table entries. Considering the detection efficiency of the DNN model, we also design some features manually in addition to the features automatically obtained by the flow table. Then we use the preprocessed data to train the DNN model and make a prediction. The overall detection framework is deployed in the SDN controller. The experiment results illustrate DNN model has higher accuracy in identifying attack traffic than machine learning algorithms, which lays a foundation for the defense against DDoS attack.

This paper mainly explores the detection and defense of DDoS attacks in the SDN architecture of the 5G environment, and proposes a DDoS attack detection method based on the deep learning two-level model CNN-LSTM in the SDN network. Not only can it greatly improve the accuracy of attack detection, but it can also reduce the time for classifying and detecting network traffic, so that the transmission of DDoS attack traffic can be blocked in time to ensure the availability of network services.

A distributed denial-of-service (DDoS) is a malicious attempt by attackers to disrupt the normal traffic of a targeted server, service or network. This is done by overwhelming the target and its surrounding infrastructure with a flood of Internet traffic. The multiple compromised computer systems (bots or zombies) then act as sources of attack traffic. Exploited machines can include computers and other network resources such as IoT devices. The attack results in either degraded network performance or a total service outage of critical infrastructure. This can lead to heavy financial losses and reputational damage. These attacks maximise effectiveness by controlling the affected systems remotely and establishing a network of bots called bot networks. It is very difficult to separate the attack traffic from normal traffic. Early detection is essential for successful mitigation of the attack, which gives rise to a very important role in cybersecurity to detect the attacks and mitigate the effects. This can be done by deploying machine learning or deep learning models to monitor the traffic data. We propose using various machine learning and deep learning algorithms to analyse the traffic patterns and separate malicious traffic from normal traffic. Two suitable datasets have been identified (DDoS attack SDN dataset and CICDDoS2019 dataset). All essential preprocessing is performed on both datasets. Feature selection is also performed before detection techniques are applied. 8 different Neural Networks/ Ensemble/ Machine Learning models are chosen and the datasets are analysed. The best model is chosen based on the performance metrics (DEEP NEURAL NETWORK MODEL). An alternative is also suggested (Next best - Hypermodel). Optimisation by Hyperparameter tuning further enhances the accuracy. Based on the nature of the attack and the intended target, suitable mitigation procedures can then be deployed.

Side Channel Attacks (SCAs), an attack that exploits the physical information generated when an encryption algorithm is executed on a device to recover the key, has become one of the key threats to the security of encrypted devices. Recently, with the development of deep learning, deep learning techniques have been applied to SCAs with good results on publicly available dataset experiences. In this paper, we propose a power traces decomposition method that divides the original power traces into two parts, where the data-influenced part is defined as data power traces (Tdata) and the other part is defined as device constant power traces, and use the Tdata for training the network model, which has more obvious advantages than using the original power traces for training the network model. To verify the effectiveness of the approach, we evaluated the ATXmega128D4 microcontroller by capturing the power traces generated when implementing AES-128. Experimental results show that network models trained using Tdata outperform network models trained using raw power traces (Traw ) in terms of classification accuracy, training time, cross-subkey recovery key, and cross-device recovery key.

Model compression is one of the most preferred techniques for efficiently deploying deep neural networks (DNNs) on resource- constrained Internet of Things (IoT) platforms. However, the simply compressed model is often vulnerable to adversarial attacks, leading to a conflict between robustness and efficiency, especially for IoT devices exposed to complex real-world scenarios. We, for the first time, address this problem by developing a novel framework dubbed Magical-Decomposition to simultaneously enhance both robustness and efficiency for hardware. By leveraging a hardware-friendly model compression method called singular value decomposition, the defending algorithm can be supported by most of the existing DNN hardware accelerators. To step further, by using a recently developed DNN interpretation tool, the underlying scheme of how the adversarial accuracy can be increased in the compressed model is highlighted clearly. Ablation studies and extensive experiments under various attacks/models/datasets consistently validate the effectiveness and scalability of the proposed framework.

Deep learning models rely on single word features and location features of text to achieve good results in text relation extraction tasks. However, previous studies have failed to make full use of semantic information contained in sentence dependency syntax trees, and data sparseness and noise propagation still affect classification models. The BERT(Bidirectional Encoder Representations from Transformers) pretrained language model provides a better representation of natural language processing tasks. And entity enhancement methods have been proved to be effective in relation extraction tasks. Therefore, this paper proposes a combination of the shortest dependency path and entity-enhanced BERT pre-training language model for model construction to reduce the impact of noise terms on the classification model and obtain more semantically expressive feature representation. The algorithm is tested on SemEval-2010 Task 8 English relation extraction dataset, and the F1 value of the final experiment can reach 0. 881.

Onion Routing is an encrypted communication system developed by the U.S. Naval Laboratory that uses existing Internet equipment to communicate anonymously. Miscreants use this means to conduct illegal transactions in the dark web, posing a security risk to citizens and the country. For this means of anonymous communication, website fingerprinting methods have been used in existing studies. These methods often have high overhead and need to run on devices with high performance, which makes the method inflexible. In this paper, we propose a lightweight method to address the high overhead problem that deep learning website fingerprinting methods generally have, so that the method can be applied on common devices while also ensuring accuracy to a certain extent. The proposed method refers to the structure of Inception net, divides the original larger convolutional kernels into smaller ones, and uses group convolution to reduce the website fingerprinting and computation to a certain extent without causing too much negative impact on the accuracy. The method was experimented on the data set collected by Rimmer et al. to ensure the effectiveness.

This study aims to explore the security issues and computational intelligence of drone information system based on deep learning. Targeting at the security issues of the drone system when it is attacked, this study adopts the improved long short-term memory (LSTM) network to analyze the cyber physical system (CPS) data for prediction from the perspective of predicting the control signal data of the system before the attack occurs. At the same time, the differential privacy frequent subgraph (DPFS) is introduced to keep data privacy confidential, and the digital twins technology is used to map the operating environment of the drone in the physical space, and an attack prediction model for drone digital twins CPS is constructed based on differential privacy-improved LSTM. Finally, the tennessee eastman (TE) process is undertaken as a simulation platform to simulate the constructed model so as to verify its performance. In addition, the proposed model is compared with the Bidirectional LSTM (BiLSTM) and Attention-BiLSTM models proposed by other scholars. It was found that the root mean square error (RMSE) of the proposed model is the smallest (0.20) when the number of hidden layer nodes is 26. Comparison with the actual flow value shows that the proposed algorithm is more accurate with better fitting. Therefore, the constructed drone attack prediction model can achieve higher prediction accuracy and obvious better robustness under the premise of ensuring errors, which can provide experimental basis for the later security and intelligent development of drone system.

Recent advances in deep learning typically, with the introduction of transformer based models has shown massive improvement and success in many Natural Language Processing (NLP) tasks. One such area which has leveraged immensely is conversational agents or chatbots in open-ended (chit-chat conversations) and task-specific (such as medical or legal dialogue bots etc.) domains. However, in the era of automation, there is still a dearth of works focused on one of the most relevant use cases, i.e., tutoring dialog systems that can help students learn new subjects or topics of their interest. Most of the previous works in this domain are either rule based systems which require a lot of manual efforts or are based on multiple choice type factual questions. In this paper, we propose EDICA (Educational Domain Infused Conversational Agent), a language tutoring Virtual Agent (VA). EDICA employs two mechanisms in order to converse fluently with a student/user over a question and assist them to learn a language: (i) Student/Tutor Intent Classification (SIC-TIC) framework to identify the intent of the student and decide the action of the VA, respectively, in the on-going conversation and (ii) Tutor Response Generation (TRG) framework to generate domain infused and intent/action conditioned tutor responses at every step of the conversation. The VA is able to provide hints, ask questions and correct student's reply by generating an appropriate, informative and relevant tutor response. We establish the superiority of our proposed approach on various evaluation metrics over other baselines and state of the art models.

Modern cyber-physical systems that comprise controlled power electronics are becoming more internet-of-things-enabled and vulnerable to cyber-attacks. Therefore, hardening those systems against cyber-attacks becomes an emerging need. In this paper, a model-based deep learning cyber-attack detection to protect electric drive systems from cyber-attacks on the physical level is proposed. The approach combines the model physics with a deep learning-based classifier. The combination of model-based and deep learning will enable more accurate cyber-attack detection results. The proposed cyber-attack detector will be trained and simulated on a PM based electric drive system to detect false data injection attacks on the drive controller command and sensor signals.

For a long time, online attacks were regarded to pose a severe threat to web - based applications, websites, and clients. It can bypass authentication methods, steal sensitive information from datasets and clients, and also gain ultimate authority of servers. A variety of ways for safeguarding online apps have been developed and used to deal the website risks. Based on the studies about the intersection of cybersecurity and machine learning, countermeasures for identifying typical web assaults have recently been presented (ML). In order to establish a better understanding on this essential topic, it is necessary to study ML methodologies, feature extraction techniques, evaluate datasets, and performance metrics utilised in a systematic manner. In this paper, we go through web security flaws like SQLi, XSS, malicious URLs, phishing attacks, path traversal, and CMDi in detail. We also go through the existing security methods for detecting these threats using machine learning approaches for URL classification. Finally, we discuss potential research opportunities for ML and DL-based techniques in this category, based on a thorough examination of existing solutions in the literature.

In recent years, as an important part of the Internet, web applications have gradually penetrated into life. Now enterprises, units and institutions are using web applications regardless of size. Intrusion detection to effectively identify malicious traffic has become an inevitable requirement for the development of network security technology. In addition, the proportion of deserialization vulnerabilities is increasing. Traditional intrusion detection mostly focuses on the identification of SQL injection, XSS, and command execution, and there are few studies on the identification of deserialization attack traffic. This paper use a method to extracts relevant features from the deserialized traffic or even the obfuscated deserialized traffic by reorganizing the traffic and running the relevant content through simulation, and combines deep learning technology to make judgments to efficiently identify deserialization attacks. Finally, a prototype system was designed to capture related attacks in real-world. The technology can be used in the field of malicious traffic detection and help combat Internet crimes in the future.

Cloud security has become a serious challenge due to increasing number of attacks day-by-day. Intrusion Detection System (IDS) requires an efficient security model for improving security in the cloud. This paper proposes a game theory based model, named as Game Theory Cloud Security Deep Neural Network (GT-CSDNN) for security in cloud. The proposed model works with the Deep Neural Network (DNN) for classification of attack and normal data. The performance of the proposed model is evaluated with CICIDS-2018 dataset. The dataset is normalized and optimal points about normal and attack data are evaluated based on the Improved Whale Algorithm (IWA). The simulation results show that the proposed model exhibits improved performance as compared with existing techniques in terms of accuracy, precision, F-score, area under the curve, False Positive Rate (FPR) and detection rate.

Aiming at the single hopping strategy in the terminal information hopping active defense technology, a variety of heterogeneous hopping modes are introduced into the terminal information hopping system, the definition of the terminal information is expanded, and the adaptive adjustment of the hopping strategy is given. A network adversarial training simulation system is researched and designed, and related subsystems are discussed from the perspective of key technologies and their implementation, including interactive adversarial training simulation system, adversarial training simulation support software system, adversarial training simulation evaluation system and adversarial training Mock Repository. The system can provide a good environment for network confrontation theory research and network confrontation training simulation, which is of great significance.

Vulnerability detection has always been an essential part of maintaining information security, and the existing work can significantly improve the performance of vulnerability detection. However, due to the differences in representation forms and deep learning models, various methods still have some limitations. In order to overcome this defect, We propose a vulnerability detection method VDBWGDL, based on weight graphs and deep learning. Firstly, it accurately locates vulnerability-sensitive keywords and generates variant codes that satisfy vulnerability trigger logic and programmer programming style through code variant methods. Then, the control flow graph is sliced for vulnerable code keywords and program critical statements. The code block is converted into a vector containing rich semantic information and input into the weight map through the deep learning model. According to specific rules, different weights are set for each node. Finally, the similarity is obtained through the similarity comparison algorithm, and the suspected vulnerability is output according to different thresholds. VDBWGDL improves the accuracy and F1 value by 3.98% and 4.85% compared with four state-of-the-art models. The experimental results prove the effectiveness of VDBWGDL.

TVM(tensor virtual machine) as a deep learning compiler which supports the conversion of machine learning models into TVM IR(intermediate representation) and to optimise the generation of high-performance machine code for various hardware platforms. While the traditional approach is to parallelise the cyclic transformations of operators, in this paper we partition the implementation of the operators in the deep learning compiler TVM with parallel scheduling to derive a faster running time solution for the operators. An optimisation algorithm for partitioning and parallel scheduling is designed for the deep learning compiler TVM, where operators such as two-dimensional convolutions are partitioned into multiple smaller implementations and several partitioned operators are run in parallel scheduling to derive the best operator partitioning and parallel scheduling decisions by means of performance estimation. To evaluate the effectiveness of the algorithm, multiple examples of the two-dimensional convolution operator, the average pooling operator, the maximum pooling operator, and the ReLU activation operator with different input sizes were tested on the CPU platform, and the performance of these operators was experimentally shown to be improved and the operators were run speedily.

Structured Query Language Injection (SQLi) is a client-side application vulnerability that allows attackers to inject malicious SQL queries with harmful intents, including stealing sensitive information, bypassing authentication, and even executing illegal operations to cause more catastrophic damage to users on the web application. According to OWASP, the top 10 harmful attacks against web applications are SQL Injection attacks. Moreover, based on data reports from the UK's National Fraud Authority, SQL Injection is responsible for 97% of data exposures. Therefore, in order to prevent the SQL Injection attack, detection SQLi system is essential. The contribution of this research is securing web applications by developing a browser extension for Google Chrome using Long Short-Term Memory (LSTM), which is a unique kind of RNN algorithm capable of learning long-term dependencies like SQL Injection attacks. The results of the model will be deployed in static analysis in a browser extension, and the LSTM algorithm will learn to identify the URL that has to be injected into Damn Vulnerable Web Application (DVWA) as a sample-tested web application. Experimental results show that the proposed SQLi detection model based on the LSTM algorithm achieves an accuracy rate of 99.97%, which means that a reliable client-side can effectively detect whether the URL being accessed contains a SQLi attack or not.

A botnet is a new type of attack method developed and integrated on the basis of traditional malicious code such as network worms and backdoor tools, and it is extremely threatening. This course combines deep learning and neural network methods in machine learning methods to detect and classify the existence of botnets. This sample does not rely on any prior features, the final multi-class classification accuracy rate is higher than 98.7%, the effect is significant.

CAPTCHA (Completely Automated Public Turing test to tell Computers and Humans Apart) is a widely used technology to distinguish real users and automated users such as bots. However, the advance of AI technologies weakens many CAPTCHA tests and can induce security concerns. In this paper, we propose a user-friendly text-based CAPTCHA generation method named Robust Text CAPTCHA (RTC). At the first stage, the foregrounds and backgrounds are constructed with font and background images respectively sampled from font and image libraries, and they are then synthesized into identifiable pseudo adversarial CAPTCHAs. At the second stage, we utilize a highly transferable adversarial attack designed for text CAPTCHAs to better obstruct CAPTCHA solvers. Our experiments cover comprehensive models including shallow models such as KNN, SVM and random forest, as well as various deep neural networks and OCR models. Experiments show that our CAPTCHAs have a failure rate lower than one millionth in general and high usability. They are also robust against various defensive techniques that attackers may employ, including adversarially trained CAPTCHA solvers and solvers trained with collected RTCs using manual annotation. Codes available at https://github.com/RulinShao/RTC.