Biblio
The rapidly growing body of research in adversarial machine learning has demonstrated that deep neural networks (DNNs) are highly vulnerable to adversarially generated images. This underscores the urgent need for practical defense techniques that can be readily deployed to combat attacks in real-time. Observing that many attack strategies aim to perturb image pixels in ways that are visually imperceptible, we place JPEG compression at the core of our proposed SHIELD defense framework, utilizing its capability to effectively "compress away" such pixel manipulation. To immunize a DNN model from artifacts introduced by compression, SHIELD "vaccinates" the model by retraining it with compressed images, where different compression levels are applied to generate multiple vaccinated models that are ultimately used together in an ensemble defense. On top of that, SHIELD adds an additional layer of protection by employing randomization at test time that compresses different regions of an image using random compression levels, making it harder for an adversary to estimate the transformation performed. This novel combination of vaccination, ensembling, and randomization makes SHIELD a fortified multi-pronged defense. We conducted extensive, large-scale experiments using the ImageNet dataset, and show that our approaches eliminate up to 98% of gray-box attacks delivered by strong adversarial techniques such as Carlini-Wagner's L2 attack and DeepFool. Our approaches are fast and work without requiring knowledge about the model.
Deep neural networks (DNNs) are known vulnerable to adversarial attacks. That is, adversarial examples, obtained by adding delicately crafted distortions onto original legal inputs, can mislead a DNN to classify them as any target labels. In a successful adversarial attack, the targeted mis-classification should be achieved with the minimal distortion added. In the literature, the added distortions are usually measured by \$L\_0\$, \$L\_1\$, \$L\_2\$, and \$L\_$\backslash$infty \$ norms, namely, L\_0, L\_1, L\_2, and L\_$ınfty$ attacks, respectively. However, there lacks a versatile framework for all types of adversarial attacks. This work for the first time unifies the methods of generating adversarial examples by leveraging ADMM (Alternating Direction Method of Multipliers), an operator splitting optimization approach, such that \$L\_0\$, \$L\_1\$, \$L\_2\$, and \$L\_$\backslash$infty \$ attacks can be effectively implemented by this general framework with little modifications. Comparing with the state-of-the-art attacks in each category, our ADMM-based attacks are so far the strongest, achieving both the 100% attack success rate and the minimal distortion.
Nowadays, network is one of the essential parts of life, and lots of primary activities are performed by using the network. Also, network security plays an important role in the administrator and monitors the operation of the system. The intrusion detection system (IDS) is a crucial module to detect and defend against the malicious traffics before the system is affected. This system can extract the information from the network system and quickly indicate the reaction which provides real-time protection for the protected system. However, detecting malicious traffics is very complicating because of their large quantity and variants. Also, the accuracy of detection and execution time are the challenges of some detection methods. In this paper, we propose an IDS platform based on convolutional neural network (CNN) called IDS-CNN to detect DoS attack. Experimental results show that our CNN based DoS detection obtains high accuracy at most 99.87%. Moreover, comparisons with other machine learning techniques including KNN, SVM, and Naïve Bayes demonstrate that our proposed method outperforms traditional ones.
This paper proposes a deep learning based method for efficient malware classification. Specially, we convert the malware classification problem into the image classification problem, which can be addressed through leveraging convolutional neural networks (CNNs). For many malware families, the images belonging to the same family have similar contours and textures, so we convert the Binary files of malware samples to uncompressed gray-scale images which possess complete information of the original malware without artificial feature extraction. We then design classifier based on Tensorflow framework of Google by combining the deep learning (DL) and malware detection technology. Experimental results show that the uncompressed gray-scale images of the malware are relatively easy to distinguish and the CNN based classifier can achieve a high success rate of 98.2%
In this study, we have used the Image Similarity technique to detect the unknown or new type of malware using CNN ap- proach. CNN was investigated and tested with three types of datasets i.e. one from Vision Research Lab, which contains 9458 gray-scale images that have been extracted from the same number of malware samples that come from 25 differ- ent malware families, and second was benign dataset which contained 3000 different kinds of benign software. Benign dataset and dataset vision research lab were initially exe- cutable files which were converted in to binary code and then converted in to image files. We obtained a testing ac- curacy of 98% on Vision Research dataset.
Deep neural network based steganalysis has developed rapidly in recent years, which poses a challenge to the security of steganography. However, there is no steganography method that can effectively resist the neural networks for steganalysis at present. In this paper, we propose a new strategy that constructs enhanced covers against neural networks with the technique of adversarial examples. The enhanced covers and their corresponding stegos are most likely to be judged as covers by the networks. Besides, we use both deep neural network based steganalysis and high-dimensional feature classifiers to evaluate the performance of steganography and propose a new comprehensive security criterion. We also make a tradeoff between the two analysis systems and improve the comprehensive security. The effectiveness of the proposed scheme is verified with the evidence obtained from the experiments on the BOSSbase using the steganography algorithm of WOW and popular steganalyzers with rich models and three state-of-the-art neural networks.
Deep learning technologies, which are the key components of state-of-the-art Artificial Intelligence (AI) services, have shown great success in providing human-level capabilities for a variety of tasks, such as visual analysis, speech recognition, and natural language processing and etc. Building a production-level deep learning model is a non-trivial task, which requires a large amount of training data, powerful computing resources, and human expertises. Therefore, illegitimate reproducing, distribution, and the derivation of proprietary deep learning models can lead to copyright infringement and economic harm to model creators. Therefore, it is essential to devise a technique to protect the intellectual property of deep learning models and enable external verification of the model ownership. In this paper, we generalize the "digital watermarking'' concept from multimedia ownership verification to deep neural network (DNNs) models. We investigate three DNN-applicable watermark generation algorithms, propose a watermark implanting approach to infuse watermark into deep learning models, and design a remote verification mechanism to determine the model ownership. By extending the intrinsic generalization and memorization capabilities of deep neural networks, we enable the models to learn specially crafted watermarks at training and activate with pre-specified predictions when observing the watermark patterns at inference. We evaluate our approach with two image recognition benchmark datasets. Our framework accurately (100$\backslash$%) and quickly verifies the ownership of all the remotely deployed deep learning models without affecting the model accuracy for normal input data. In addition, the embedded watermarks in DNN models are robust and resilient to different counter-watermark mechanisms, such as fine-tuning, parameter pruning, and model inversion attacks.
Automated network control and management has been a long standing target of network protocols. We address in this paper the question of automated protocol design, where distributed networked nodes have to cooperate to achieve a common goal without a priori knowledge on which information to exchange or the network topology. While reinforcement learning has often been proposed for this task, we propose here to apply recent methods from semi-supervised deep neural networks which are focused on graphs. Our main contribution is an approach for applying graph-based deep learning on distributed routing protocols via a novel neural network architecture named Graph-Query Neural Network. We apply our approach to the tasks of shortest path and max-min routing. We evaluate the learned protocols in cold-start and also in case of topology changes. Numerical results show that our approach is able to automatically develop efficient routing protocols for those two use-cases with accuracies larger than 95%. We also show that specific properties of network protocols, such as resilience to packet loss, can be explicitly included in the learned protocol.
Silicon Physical Unclonable Function (PUF) is arguably the most promising hardware security primitive. In particular, PUFs that are capable of generating a large amount of challenge response pairs (CRPs) can be used in many security applications. However, these CRPs can also be exploited by machine learning attacks to model the PUF and predict its response. In this paper, we first show that, based on data in the public domain, two popular PUFs that can generate CRPs (i.e., arbiter PUF and reconfigurable ring oscillator (RO) PUF) can be broken by simple logistic regression (LR) attack with about 99% accuracy. We then propose a feedback structure to XOR the PUF response with the challenge and challenge the PUF again to generate the response. Results show that this successfully reduces LR's learning accuracy to the lower 50%, but artificial neural network (ANN) learning attack still has an 80% success rate. Therefore, we propose a configurable ring oscillator based dual-mode PUF which works with both odd number of inverters (like the reconfigurable RO PUF) and even number of inverters (like a bistable ring (BR) PUF). Since currently there are no known attacks that can model both RO PUF and BR PUF, the dual-mode PUF will be resistant to modeling attacks as long as we can hide its working mode from the attackers, which we achieve with two practical methods. Finally, we implement the proposed dual-mode PUF on Nexys 4 FPGA boards and collect real measurement to show that it reduces the learning accuracy of LR and ANN to the mid-50% and low 60%, respectively. In addition, it meets the PUF requirements of uniqueness, randomness, and robustness.
The power of artificial neural networks to form predictive models for phenomenon that exhibit non-linear relationships is a given fact. Despite this advantage, artificial neural networks are known to suffer drawbacks such as long training times and computational intensity. The researchers propose a two-tiered approach to enhance the learning performance of artificial neural networks for phenomenon with time series where data exhibits predictable changes that occur every calendar year. This paper focuses on the initial results of the first phase of the proposed algorithm which incorporates clustering and classification prior to application of the backpropagation algorithm. The 2016–2017 zonal load data of France is used as the data set. K-means is chosen as the clustering algorithm and a comparison is made between Naïve Bayes and k-Nearest Neighbors to determine the better classifier for this data set. The initial results show that electrical load behavior is not necessarily reflective of calendar clustering even without using the min-max temperature recorded during the inclusive months. Simulating the day-type classification process using one cluster, initial results show that the k-nearest neighbors outperforms the Naïve Bayes classifier for this data set and that the best feature to be used for classification into day type is the daily min-max load. These classified load data is expected to reduce training time and improve the overall performance of short-term load demand predictive models in a future paper.
With an increase in targeted attacks such as advanced persistent threats (APTs), enterprise system defenders require comprehensive frameworks that allow them to collaborate and evaluate their defense systems against such attacks. MITRE has developed a framework which includes a database of different kill-chains, tactics, techniques, and procedures that attackers employ to perform these attacks. In this work, we leverage natural language processing techniques to extract attacker actions from threat report documents generated by different organizations and automatically classify them into standardized tactics and techniques, while providing relevant mitigation advisories for each attack. A naïve method to achieve this is by training a machine learning model to predict labels that associate the reports with relevant categories. In practice, however, sufficient labeled data for model training is not always readily available, so that training and test data come from different sources, resulting in bias. A naïve model would typically underperform in such a situation. We address this major challenge by incorporating an importance weighting scheme called bias correction that efficiently utilizes available labeled data, given threat reports, whose categories are to be automatically predicted. We empirically evaluated our approach on 18,257 real-world threat reports generated between year 2000 and 2018 from various computer security organizations to demonstrate its superiority by comparing its performance with an existing approach.
The current AI revolution provides us with many new, but often very complex algorithmic systems. This complexity does not only limit understanding, but also acceptance of e.g. deep learning methods. In recent years, explainable AI (XAI) has been proposed as a remedy. However, this research is rarely supported by publications on explanations from social sciences. We suggest a bottom-up approach to explanations for (game) AI, by starting from a baseline definition of understandability informed by the concept of limited human working memory. We detail our approach and demonstrate its application to two games from the GVGAI framework. Finally, we discuss our vision of how additional concepts from social sciences can be integrated into our proposed approach and how the results can be generalised.
In this paper, we present a framework for graph-based representation of relation between sensors and alert types in a security alert sharing platform. Nodes in a graph represent either sensors or alert types, while edges represent various relations between them, such as common type of reported alerts or duplicated alerts. The graph is automatically updated, stored in a graph database, and visualized. The resulting graph will be used by network administrators and security analysts as a visual guide and situational awareness tool in a complex environment of security alert sharing.
In an Internet of Things (IOT) network, each node (device) provides and requires services and with the growth in IOT, the number of nodes providing the same service have also increased, thus creating a problem of selecting one reliable service from among many providers. In this paper, we propose a scalable graph-based collaborative filtering recommendation algorithm, improved using trust to solve service selection problem, which can scale to match the growth in IOT unlike a central recommender which fails. Using this recommender, a node can predict its ratings for the nodes that are providing the required service and then select the best rated service provider.
Policies govern choices in the behavior of systems. They are applied to human behavior as well as to the behavior of autonomous systems but are defined differently in each case. Generally humans have the ability to interpret the intent behind the policies, to bring about their desired effects, even occasionally violating them when the need arises. In contrast, policies for automated systems fully define the prescribed behavior without ambiguity, conflicts or omissions. The increasing use of AI techniques and machine learning in autonomous systems such as drones promises to blur these boundaries and allows us to conceive in a similar way more flexible policies for the spectrum of human-autonomous systems collaborations. In coalition environments this spectrum extends across the boundaries of authority in pursuit of a common coalition goal and covers collaborations between human and autonomous systems alike. In social sciences, social exchange theory has been applied successfully to explain human behavior in a variety of contexts. It provides a framework linking the expected rewards, costs, satisfaction and commitment to explain and anticipate the choices that individuals make when confronted with various options. We discuss here how it can be used within coalition environments to explain joint decision making and to help formulate policies re-framing the concepts where appropriate. Social exchange theory is particularly attractive within this context as it provides a theory with “measurable” components that can be readily integrated in machine reasoning processes.
Current technologies to include cloud computing, social networking, mobile applications and crowd and synthetic intelligence, coupled with the explosion in storage and processing power, are evolving massive-scale marketplaces for a wide variety of resources and services. They are also enabling unprecedented forms and levels of collaborations among human and machine entities. In this new era, trust remains the keystone of success in any relationship between two or more parties. A primary challenge is to establish and manage trust in environments where massive numbers of consumers, providers and brokers are largely autonomous with vastly diverse requirements, capabilities, and trust profiles. Most contemporary trust management solutions are oblivious to diversities in trustors' requirements and contexts, utilize direct or indirect experiences as the only form of trust computations, employ hardcoded trust computations and marginally consider collaboration in trust management. We surmise the need for reference architecture for trust management to guide the development of a wide spectrum of trust management systems. In our previous work, we presented a preliminary reference architecture for trust management which provides customizable and reconfigurable trust management operations to accommodate varying levels of diversity and trust personalization. In this paper, we present a comprehensive taxonomy for trust management and extend our reference architecture to feature collaboration as a first-class object. Our goal is to promote the development of new collaborative trust management systems, where various trust management operations would involve collaborating entities. Using the proposed architecture, we implemented a collaborative personalized trust management system. Simulation results demonstrate the effectiveness and efficiency of our system.
Mobile Ad Hoc Network (MANET) technology provides intercommunication between different nodes where no infrastructure is available for communication. MANET is attracting many researcher attentions as it is cost effective and easy for implementation. Main challenging aspect in MANET is its vulnerability. In MANET nodes are very much vulnerable to attacks along with its data as well as data flowing through these nodes. One of the main reasons of these vulnerabilities is its communication policy which makes nodes interdependent for interaction and data flow. This mutual trust between nodes is exploited by attackers through injecting malicious node or replicating any legitimate node in MANET. One of these attacks is blackhole attack. In this study, the behavior of blackhole attack is discussed and have proposed a lightweight solution for blackhole attack which uses inbuilt functions.