Biblio

Internal attacks are one of the biggest cybersecurity issues to companies and businesses. Despite the implemented perimeter security systems, the risk of adversely affecting the security and privacy of the organization’s information remains very high. Actually, the detection of such a threat is known to be a very complicated problem, presenting many challenges to the research community. In this paper, we investigate the effectiveness and usefulness of using Autoencoder and Variational Autoencoder deep learning algorithms to automatically defend against insider threats, without human intervention. The performance evaluation of the proposed models is done on the public CERT dataset (CERT r4.2) that contains both benign and malicious activities generated from 1000 simulated users. The comparison results with other models show that the Variational Autoencoder neural network provides the best overall performance with a higher detection accuracy and a reasonable false positive rate.

How to develop AI systems that can explain how they made decisions is one of the important and hot topics today. Inspired by the dual-process theory in psychology, this paper proposes a human-in-the-loop approach to develop System-2 AI that makes an inference logically and outputs interpretable explanation. Our proposed method first asks crowd workers to raise understandable features of objects of multiple classes and collect training data from the Internet to generate classifiers for the features. Logical decision rules with the set of generated classifiers can explain why each object is of a particular class. In our preliminary experiment, we applied our method to an image classification of Asian national flags and examined the effectiveness and issues of our method. In our future studies, we plan to combine the System-2 AI with System-1 AI (e.g., neural networks) to efficiently output decisions.

Scene text detection methods have developed greatly in the past few years. However, due to the limitation of the diversity of the text background of natural scene, the previous methods often failed when detecting more complicated text instances (e.g., super-long text and arbitrarily shaped text). In this paper, a text detection method based on multi -channel bounding box fusion is designed to address the problem. Firstly, the convolutional neural network is used as the basic network for feature extraction, including shallow text feature map and deep semantic text feature map. Secondly, the whole convolutional network is used for upsampling of feature map and fusion of feature map at each layer, so as to obtain pixel-level text and non-text classification results. Then, two independent text detection boxes channels are designed: the boundary box regression channel and get the bounding box directly on the score map channel. Finally, the result is obtained by combining multi-channel boundary box fusion mechanism with the detection box of the two channels. Experiments on ICDAR2013 and ICDAR2015 demonstrate that the proposed method achieves competitive results in scene text detection.

Hardware Trojans are modifications made by malicious insiders or third party providers during the design or fabrication phase of the IC (Integrated Circuits) design cycle in a covert manner. These cause catastrophic consequences ranging from manipulating the functionality of individual blocks to disabling the entire chip. Thus, a need for detecting trojans becomes necessary. In this work, we propose a deep learning based approach for detecting trojans in IC chips. In particular, we insert trojans at the circuit-level and generate data by measuring power during normal operation and under attack. Further, we develop deep learning models using Neural networks and Auto-encoders to analyze datasets for outlier detection by profiling the normal behavior and leveraging them to detect anomalies in power consumption. Our approach is generic and non-invasive in that it can be applied to any block without any modifications to the design. Evaluation of the proposed approach shows an accuracy ranging from 92.23% to 99.33% in detecting trojans.

Vulnerabilities in the source code of software are critical issues in the realm of software engineering. Coping with vulnerabilities in software source code is becoming more challenging due to several aspects of complexity and volume. Deep learning has gained popularity throughout the years as a means of addressing such issues. In this paper, we propose an evaluation of vulnerability detection performance on source code representations and evaluate how Machine Learning (ML) strategies can improve them. The structure of our experiment consists of 3 Deep Neural Networks (DNNs) in conjunction with five different source code representations; Abstract Syntax Trees (ASTs), Code Gadgets (CGs), Semantics-based Vulnerability Candidates (SeVCs), Lexed Code Representations (LCRs), and Composite Code Representations (CCRs). Experimental results show that employing different ML strategies in conjunction with the base model structure influences the performance results to a varying degree. However, ML-based techniques suffer from poor performance on class imbalance handling when used in conjunction with source code representations for software vulnerability detection.

Edge nodes are crucial for detection against multitudes of cyber attacks on Internet-of-Things endpoints and is set to become part of a multi-billion industry. The resource constraints in this novel network infrastructure tier constricts the deployment of existing Network Intrusion Detection System with Deep Learning models (DLM). We address this issue by developing a novel light, fast and accurate `Edge-Detect' model, which detects Distributed Denial of Service attack on edge nodes using DLM techniques. Our model can work within resource restrictions i.e. low power, memory and processing capabilities, to produce accurate results at a meaningful pace. It is built by creating layers of Long Short-Term Memory or Gated Recurrent Unit based cells, which are known for their excellent representation of sequential data. We designed a practical data science pipeline with Recurring Neural Network to learn from the network packet behavior in order to identify whether it is normal or attack-oriented. The model evaluation is from deployment on actual edge node represented by Raspberry Pi using current cybersecurity dataset (UNSW2015). Our results demonstrate that in comparison to conventional DLM techniques, our model maintains a high testing accuracy of 99% even with lower resource utilization in terms of cpu and memory. In addition, it is nearly 3 times smaller in size than the state-of-art model and yet requires a much lower testing time.

Aiming at the problems of imperfect dynamic verification of power grid security and stability control strategy and high test cost, a reliability test method of power grid security control system based on BP neural network and dynamic group simulation is proposed. Firstly, the fault simulation results of real-time digital simulation system (RTDS) software are taken as the data source, and the dynamic test data are obtained with the help of the existing dispatching data network, wireless virtual private network, global positioning system and other communication resources; Secondly, the important test items are selected through the minimum redundancy maximum correlation algorithm, and the test items are used to form a feature set, and then the BP neural network model is used to predict the test results. Finally, the dynamic remote test platform is tested by the dynamic whole group simulation of the security and stability control system. Compared with the traditional test methods, the proposed method reduces the test cost by more than 50%. Experimental results show that the proposed method can effectively complete the reliability test of power grid security control system based on dynamic group simulation, and reduce the test cost.

Recently, federated learning (FL), as an advanced and practical solution, has been applied to deal with privacy-preserving issues in distributed multi-party federated modeling. However, most existing FL methods focus on the same privacy-preserving budget while ignoring various privacy requirements of participants. In this paper, we for the first time propose an algorithm (PLU-FedOA) to optimize the deep neural network of horizontal FL with personalized local differential privacy. For such considerations, we design two approaches: PLU, which allows clients to upload local updates under differential privacy-preserving of personally selected privacy level, and FedOA, which helps the server aggregates local parameters with optimized weight in mixed privacy-preserving scenarios. Moreover, we theoretically analyze the effect on privacy and optimization of our approaches. Finally, we verify PLU-FedOA on real-world datasets.

Keystroke dynamics is a behavioural biometric form of authentication based on the inherent typing behaviour of an individual. While this technique is gaining traction, protecting the privacy of the users is of utmost importance. Fully Homomorphic Encryption is a technique that allows performing computation on encrypted data, which enables processing of sensitive data in an untrusted environment. FHE is also known to be “future-proof” since it is a lattice-based cryptosystem that is regarded as quantum-safe. It has seen significant performance improvements over the years with substantially increased developer-friendly tools. We propose a neural network for keystroke analysis trained using differential privacy to speed up training while preserving privacy and predicting on encrypted data using FHE to keep the users' privacy intact while offering sufficient usability.

Recent advances in video manipulation techniques have made the generation of fake videos more accessible than ever before. Manipulated videos can fuel disinformation and reduce trust in media. Therefore detection of fake videos has garnered immense interest in academia and industry. Recently developed Deepfake detection methods rely on Deep Neural Networks (DNNs) to distinguish AI-generated fake videos from real videos. In this work, we demonstrate that it is possible to bypass such detectors by adversarially modifying fake videos synthesized using existing Deepfake generation methods. We further demonstrate that our adversarial perturbations are robust to image and video compression codecs, making them a real-world threat. We present pipelines in both white-box and black-box attack scenarios that can fool DNN based Deepfake detectors into classifying fake videos as real.

Advances in machine learning, deep learning, and Artificial Intelligence(AI) allows people to exchange other people's faces and voices in videos to make it look like what they did or say whatever you want to say. These videos and photos are called “deepfake” and are getting more complicated every day and this has lawmakers worried. This technology uses machine learning technology to provide computers with real data about images, so that we can make forgeries. The creators of Deepfake use artificial intelligence and machine learning algorithms to mimic the work and characteristics of real humans. It differs from counterfeit traditional media because it is difficult to identify. As In the 2020 elections loomed, AI-generated deepfakes were hit the news cycle. DeepFakes threatens facial recognition and online content. This deception can be dangerous, because if used incorrectly, this technique can be abused. Fake video, voice, and audio clips can do enormous damage. This paper examines the algorithms used to generate deepfakes as well as the methods proposed to detect them. We go through the threats, research patterns, and future directions for deepfake technologies in detail. This research provides a detailed description of deep imitation technology and encourages the creation of new and more powerful methods to deal with increasingly severe deep imitation by studying the history of deep imitation.

This paper aims to address the problem of Deepfake Detection using a Two-Stream Capsule Network. First we review methods used to create Deepfake content, as well as methods proposed in the literature to detect such Deepfake content. We then propose a novel architecture to detect Deepfakes, which consists of a two-stream Capsule network running in parallel that takes in both RGB images/frames as well as Error Level Analysis images. Results show that the proposed approach exhibits the detection accuracy of 73.39 % and 57.45 % for the Deepfake Detection Challenge (DFDC) and the Celeb-DF datasets respectively. These results are, however, from a preliminary implementation of the proposed approach. As part of future work, population-based optimization techniques such as Particle Swarm Optimization (PSO) will be used to tune the hyper parameters for better performance.

Video analytics has many applications in traffic control, security monitoring, action/event analysis, etc. With the adoption of deep neural networks, the accuracy of video analytics in video streams has been greatly improved. However, deep neural networks for performing video analytics are compute-intensive. In order to reduce processing time, many systems switch to the lower frame rate or resolution. State-of-the-art switching approaches adjust configurations by profiling video clips on a large configuration space. Multiple configurations are tested periodically and the cheapest one with a desired accuracy is adopted. In this paper, we propose a method that adapts the configuration by analyzing past video analytics results instead of profiling candidate configurations. Our method adopts a lower/higher resolution or frame rate when objects move slow/fast. We train a model that automatically selects the best configuration. We evaluate our method with two real-world video analytics applications: traffic tracking and pose estimation. Compared to the periodic profiling method, our method achieves 3%-12% higher accuracy with the same resource cost and 8-17x faster with comparable accuracy.

Nowadays, video surveillance cameras are widely used in many smart city applications for ensuring road safety. We can use video data from them to solve such tasks as traffic management, driving control, environmental monitoring, etc. Most of these applications are based on object recognition and tracking algorithms. However, the video image quality is not always meet the requirements of such algorithms due to the influence of different external factors. A variety of adverse weather conditions produce noise on the images, which often makes it difficult to detect objects correctly. Lately, deep learning methods show good results in image processing, including denoising tasks. This work is devoted to the study of using these methods for image quality enhancement in difficult weather conditions such as snow, rain, fog. Different deep learning techniques were evaluated in terms of their impact on the quality of object detection/recognition. Finally, the system for automatic image denoising was developed.

For exploiting multimode fiber optic communication networks towards physical layer security, we have trained a neural network performing mode decomposition of 10 modes. The approach is based on intensity-only camera images and works in real-time.

Authorship attribution, an NLP problem where anonymous text is matched to its author, has important, cross-disciplinary applications, particularly those concerning cyber-defense. Our research examines the degree of sensitivity that attention-based models have to adversarial perturbations. We ask, what is the minimal amount of change necessary to maximally confuse a transformer model? In our investigation we examine a balanced subset of emails from the Enron email dataset, calculating the performance of our model before and after email signatures have been perturbed. Results show that the model's performance changed significantly in the absence of a signature, indicating the importance of email signatures in email authorship detection. Furthermore, we show that these models rely on signatures for shorter emails much more than for longer emails. We also indicate that additional research is necessary to investigate stylometric features and adversarial training to further improve classification model robustness.

Distributed denial of service (DDoS) attack is a common way of network attack. It has the characteristics of wide distribution, low cost and difficult defense. The traditional algorithms of machine learning (ML) have such shortcomings as excessive systemic overhead and low accuracy in detection of DDoS. In this paper, a CGAN (conditional generative adversarial networks, conditional GAN) -based method is proposed to detect the attack of DDoS. On off-line training, five features are extracted in order to adapt the input of neural network. On the online recognition, CGAN model is adopted to recognize the packets of DDoS attack. The experimental results demonstrate that our proposed method obtains the better performance than the random forest-based method.

Today cyber-attacks to critical infrastructures can perform outages, economical loss, physical damage to people and the environment, among many others. In particular, the smart grid is one of the main targets. In this paper, we develop and evaluate software detectors for integrity attacks to smart meter readings. The detectors rely upon different techniques and models, such as autoregressive models, clustering, and neural networks. Our evaluation considers different “attack scenarios”, then resembling the plethora of attacks found in last years. Starting from previous works in the literature, we carry out a detailed experimentation and analysis, so to identify which “detectors” best fit for each “attack scenario”. Our results contradict some findings of previous works and also offer a light for choosing the techniques that can address best the attacks to smart meters.

Network traffic detection is closely related to network security, and it is also a hot research topic now. With the development of encryption technology, traffic detection has become more and more difficult, and many crimes have occurred on the dark web, so how to detect dark web traffic is the subject of this study. In this paper, we proposed a dark web traffic(Tor traffic) detection scheme based on deep learning and conducted experiments on public data sets. By analyzing the results of the experiment, our detection precision rate reached 95.47%.

In contrast to traditional grids, smart grids can help utilities save energy, thereby reducing operating costs. In the smart grid, the quality of monitoring and control can be fully improved by combining computing and intelligent communication knowledge. However, this will expose the system to FDI attacks, and the system is vulnerable to intrusion. Therefore, it is very important to detect such erroneous data injection attacks and provide an algorithm to protect the system from such attacks. In this paper, a FDI detection method based on LSTM has been proposed, which is validated by the simulation on the ieee-14 bus platform.

A recent source of concern for the security of neural networks is the emergence of clean-label dataset poisoning attacks, wherein correctly labeled poison samples are injected into the training dataset. While these poison samples look legitimate to the human observer, they contain malicious characteristics that trigger a targeted misclassification during inference. We propose a scalable and transferable clean-label poisoning attack against transfer learning, which creates poison images with their center close to the target image in the feature space. Our attack, Bullseye Polytope, improves the attack success rate of the current state-of-the-art by 26.75% in end-to-end transfer learning, while increasing attack speed by a factor of 12. We further extend Bullseye Polytope to a more practical attack model by including multiple images of the same object (e.g., from different angles) when crafting the poison samples. We demonstrate that this extension improves attack transferability by over 16% to unseen images (of the same object) without using extra poison samples.

Despite the functional success of deep neural networks (DNNs), their trustworthiness remains a crucial open challenge. To address this challenge, both testing and verification techniques have been proposed. But these existing techniques pro- vide either scalability to large networks or formal guarantees, not both. In this paper, we propose a scalable quantitative verification framework for deep neural networks, i.e., a test-driven approach that comes with formal guarantees that a desired probabilistic property is satisfied. Our technique performs enough tests until soundness of a formal probabilistic property can be proven. It can be used to certify properties of both deterministic and randomized DNNs. We implement our approach in a tool called PROVERO1 and apply it in the context of certifying adversarial robustness of DNNs. In this context, we first show a new attack- agnostic measure of robustness which offers an alternative to purely attack-based methodology of evaluating robustness being reported today. Second, PROVERO provides certificates of robustness for large DNNs, where existing state-of-the-art verification tools fail to produce conclusive results. Our work paves the way forward for verifying properties of distributions captured by real-world deep neural networks, with provable guarantees, even where testers only have black-box access to the neural network.

Style transfer means using a neural network to extract the content of one image and the style of the other image. The two are combined to get the final result, broadly applied in social communication, animation production, entertainment items. Using style transfer, users can share and exchange images; painters can create specific art styles more readily with less creation cost and production time. Therefore, style transfer is widely concerned recently due to its various and valuable applications. In the past few years, the paper reviews style transfer and chooses three representative works to analyze in detail and contrast with each other, including StyleGAN, CycleGAN, and TL-GAN. Moreover, what function an ideal model of style transfer should realize is discussed. Compared with such a model, potential problems and prospects of different methods to achieve style transfer are listed. A couple of solutions to these drawbacks are given in the end.

Arbitrary style transfer refers to generate a new image based on any set of existing images. Meanwhile, the generated image retains the content structure of one and the style pattern of another. In terms of content retention and style transfer, the recent arbitrary style transfer algorithms normally perform well in one, but it is difficult to find a trade-off between the two. In this paper, we propose the Deep Content Guidance Network (DCGN) which is stacked by content guidance (CG) layers. And each CG layer involves one position self-attention (pSA) module, one channel self-attention (cSA) module and one content guidance attention (cGA) module. Specially, the pSA module extracts more effective content information on the spatial layout of content images and the cSA module makes the style representation of style images in the channel dimension richer. And in the non-local view, the cGA module utilizes content information to guide the distribution of style features, which obtains a more detailed style expression. Moreover, we introduce a new permutation loss to generalize feature expression, so as to obtain abundant feature expressions while maintaining content structure. Qualitative and quantitative experiments verify that our approach can transform into better stylized images than the state-of-the-art methods.

In recent years, the application of style transfer has become more and more widespread. Traditional deep learning-based style transfer networks often have problems such as image distortion, loss of detailed information, partial content disappearance, and transfer errors. The style transfer network based on deep learning that we propose in this article is aimed at dealing with these problems. Our method uses image edge information fusion and semantic segmentation technology to constrain the image structure before and after the migration, so that the converted image maintains structural consistency and integrity. We have verified that this method can successfully suppress image conversion distortion in most scenarios, and can generate good results.