Search

Deep neural networks (DNN) are producing groundbreaking results in virtually all academic and commercial domains and will serve as the workhorse of future human-machine teams that will modernize the Department of Defense (DOD). As such, leaders will need to trust and rely on these networks, which makes their security a paramount concern. Considerable research has demonstrated that DNNs remain vulnerable to adversarial examples. While many defense schemes have been proposed to counter the equally many attack vectors, none have been successful at securing a DNN from this vulnerability. Novel attacks expose blind spots unique to a network’s defense, indicating the need for a robust and adaptable attack, used to expose these vulnerabilities early in the development phase. We propose a novel reinforcement learning–based attack, Adversarial Reinforcement Learning Agent (ARLA), designed to learn the vulnerabilities of a DNN and generate adversarial examples to exploit them. ARLA was able to significantly degrade the accuracy of five CIFAR-10 DNNs, four of which used a state-of-the-art defense. We compared our method to other state-of-the-art attacks and found evidence that ARLA is an adaptive attack, making it a useful tool for testing the reliability of DNNs before they are deployed within the DOD., Lieutenant Commander, United States Navy, Approved for public release. Distribution is unlimited.

The Onion Router (Tor) is used by adversaries and warfighters alike to encrypt session information and gain anonymity on the internet. Since its creation in 2002, Tor has gained popularity by terrorist organizations, human traffickers, and illegal drug distributors who wish to use Tor services to mask their identity while engaging in illegal activities. Fingerprinting attacks assist in thwarting these attempts. Website fingerprinting (WF) attacks have been proven successful at linking a user to the website they have viewed over an encrypted Tor connection. With consumer video streaming traffic making up a large majority of internet traffic and sites like YouTube remaining in the top visited sites in the world, it is just as likely that adversaries are using videos to spread misinformation, illegal content, and terrorist propaganda. Video fingerprinting (VF) attacks look to use encrypted network traffic to predict the content of encrypted video sessions in closed- and open-world scenarios. This research builds upon an existing dataset of encrypted video session data and use statistical analysis to train a machine-learning classifier, using deep fingerprinting (DF), to predict videos viewed over Tor. DF is a machine learning technique that relies on the use of convolutional neural networks (CNN) and can be used to conduct VF attacks against Tor. By analyzing the results of these experiments, we can more accurately identify malicious video streaming activity over Tor., Civilian, Approved for public release. Distribution is unlimited.

Neural networks (NNs) are vulnerable to adversarial examples, and extensive research is aimed at detecting them. However, detecting adversarial examples is not easy, even with the construction of new loss functions in a network. In this study, we introduce the Adversarial Mix up (AdvMix) network, a neural network that adds a None of the Above (NOTA) class on top of the existing classes to isolate the space where adversarial examples exist. We investigate the effectiveness of AdvMix in improving the robustness of models trained on deep neural networks against adversarial attacks by detecting them. We experimented with various data augmentation techniques and trained nine different models. Our findings show that using an AdvMix network can significantly improve the performance of models against various attacks while achieving better accuracy on benign examples. We were able to increase the accuracy of the vanilla model from 91% to 95% and improve the model's robustness. In many cases, we were able to eliminate the vulnerability of models against some popular and efficient attacks., Captain, Hellenic Army, Approved for public release. Distribution is unlimited.

Machine learning is found in nearly every facet of daily life. Large amounts of data are required but not always available for specific problems, precluding the use of advanced methods such as deep learning and convolutional neural networks. The Euclidean Network (EN) can be used to mitigate these issues. The EN was thoroughly tested to prove its viability as a classification algorithm and that its methods may be used to augment data and transform the input data to increase its feature space dimensionality. Originally, it was hypothesized that the EN could be used to synthetically generate data to augment a data set, though this method was proven to be ineffective. The next area of research sought to expand the dimensionality of the input feature space to improve performance with additional classifiers. This area showed positive results, which supported the hypothesis that more complex, dense input would give algorithms more insight into the data and improve performance. The EN has been found to perform exceptionally well as an independent classifier, as it achieved the highest accuracy for 12 of the 21 data sets. For the remaining 9, though it did not have the highest accuracy, the EN performed comparably to more sophisticated algorithms. The EN also proved capable to expand a data set's feature space to further improve performance. This tactic provided a more robust classification technique and saw an average increase in accuracy of 3% between all data sets., Captain, United States Marine Corps, Approved for public release. Distribution is unlimited.

Firewalls are key for maintaining a secure network, but it cannot be assumed that network traffic that manages to get through one is completely safe. Anomaly detection refers to methods that can be used to discover unique or uncommon occurrences within a particular dataset. Unsupervised machine learning techniques involve machine learning with unlabeled data, and can be utilized in order to perform anomaly detection by ingesting a given set of data and finding instances that diverge from the rest in meaningful ways that may not be obvious to the human eye. In this study we aim to analyze anomalies that are detected in incoming packet and flow network traffic data that successfully passed through a firewall and determine what significance there may be within such anomalies. Considering the vast amount of malicious traffic that exists and gets generated regularly, this study shows that Gaussian Mixtures can be used for discovery of anomalies within network traffic that passed through a firewall to discover potential undesirable or malicious traffic., Civilian, Approved for public release. Distribution is unlimited.

Although federated learning and Bayesian neural networks have been researched, there are few implementations of the federated learning of Bayesian networks. In this thesis, a federated learning training environment for Bayesian neural networks using a public code base, Flower, is developed. With it is the exploration of state-of-the-art architecture, residual networks, and Bayesian versions of it. These architectures are then tested with independently and identically distributed (IID) datasets and non-IID datasets derived from the Dirichlet distribution. Results show that the MC Dropout version of Bayesian neural networks can achieve state-of-the-art results—91% accuracy—for IID partitions of the CIFAR10 dataset through federated learning. When the partitions are non-IID, federated learning through inverse variance aggregation of probabilistic weights does as well as its deterministic counterpart, with roughly 83% accuracy. This shows that Bayesian neural networks can be federated and achieve state-of-the-art results as well., Outstanding Thesis, Lieutenant, United States Navy, Approved for public release. Distribution is unlimited.

Adversarial example attacks alter an image so the image appears largely unaltered to human eyes, but image-recognition models will misclassify it. This is a common type of attack, against which there is currently no good general defense. Most state-of-the-art methods of detecting adversarial example attacks only consistently succeed in recognizing a few known attacks. These defenses do not generalize well to detecting other attacks, which means an adversary only needs to change their attack to leave us without robust abilities to detect attacks. Military intelligence increasingly relies on machine learning image recognition for analyzing satellite images. Finding defenses against these adversarial example attacks is important for ensuring our intelligence-gathering capabilities are not compromised. This thesis seeks to contribute models which will push the state of the art towards successful recognition of adversarial attacks regardless of which type of attack was used. Models we named 3-Mix were trained using combinations of different attacked images; other models were trained using SaliencyMix. These defenses were evaluated against ten attacks: PGD, auto-PGD, autoattack, square, Carlini L2 and L-inf, deepfool, elasticnet, JSMA, and boundary. On average the attack success rate against the best defense model was 0.12 for 3-Mix, 0.31 for SaliencyMix, and 0.77 for comparison model Mixup., Civilian, Copyright is reserved by the copyright owner.

Detecting malicious traffic on networks is a critical problem facing the Department of Defense. In this thesis we utilize cutting edge machine learning techniques to detect malicious network traffic. We begin with two real-world datasets. First, real internet traffic collected on the NPS Enterprise Research Network, and second, the IoT23 dataset consisting of internet of things (IoT) devices that have been infected with malware. We convert raw packet captures into TCP streams using the 5-tuple definition from RFC6146. We then apply the traffic interaction graph (TIG) framework to these flows to capture burst patterns among signed packet lengths. Finally, we train these flows on a random forest classifier (RFC), a simple convolutional neural network (CNN), and a graph convolutional network (GCN). Additionally, we simulate various attack levels by combining the two datasets at various levels (99% NPS to 1% IoT, 99-5, 90-10, and IoT23 only). In each of these models we get exceptional results in accuracy, precision, and recall. This work specifically provides a proof of concept for using the TIG framework and graph neural networks to classify TCP flows. Future work should explore model enhancement, data enrichment, or stream-lining the entire process into a real-time software package., Outstanding Thesis, Lieutenant, United States Navy, Approved for public release. Distribution is unlimited.

Deception methods have been applied to the traditional domains of war (air, land, sea, and space). In the newest domain of cyber, deception can be studied to see how it can be best used. Cyberspace operations are an essential warfighting domain within the Department of Defense (DOD). Many training exercises and courses have been developed to aid leadership with planning and to execute cyberspace effects that support operations. However, only a few simulations train cyber operators about how to respond to cyberspace threats. This work tested a commercial product from Soar Technologies (Soar Tech) that simulates conflict in cyberspace. The Cyberspace Course of Action Tool (CCAT) is a decision-support tool that evaluates defensive deception in a wargame simulating a local-area network being attacked. Results showed that defensive deception methods of decoys and bait could be effective in cyberspace. This could help military cyber defenses since their digital infrastructure is threatened daily with cyberattacks.

Deep learning is becoming a technology central to the safety and accuracy of many types of systems. Unfortunately, attackers can create adversarial examples that manipulate Deep Neural Networks (DNN) into making incorrect predictions by carefully crafting perturbations that, to humans, look indistinguishable from examples the DNN would classify correctly. Research shows that adversarial examples exist near the decision boundary. Decision-based attacks are designed to find adversarial examples by traversing the data manifold toward the decision boundary using iterative sampling without any knowledge of the model parameters or gradients. In this sense, decision-based attacks are very important, as they apply to many real-world attack scenarios. We propose a new decision-based attack, High Entropy Input Sampling (HEIS), that iteratively steps toward the decision boundary by using entropy over class predictions as a heuristic to find adversarial examples without any knowledge of the model gradients. Using HEIS, we were able to produce adversarial examples that reduced the accuracy of a CIFAR-10 DNN from 91% to 11% for epsilon=0.2 and reduced the accuracy of ResNet50, an ImageNet DNN, from 81% to 22% for epsilon=0.4. Furthermore, we discovered that the adversarial examples are highly transferable to other models, causing dramatic drops in accuracy among all models tested. Finally, we use HEIS to break three state-of-the-art neural network defenses., Lieutenant, United States Navy, Approved for public release. Distribution is unlimited.

The future battlefield is one that will be heavily influenced by rapidly changing technological capabilities of near-peer adversaries. Success in this environment will require simple to use systems that are adaptable to situations and capable of integration with other forces and systems. The Multi-Domain Operations Command, Control, Computers, Communications, Combat Systems, and Intelligence (MDOC5i) aims to prepare the Marine Corps for the future battlefield. Due to certain drawback of traditional machine learning techniques, MDOC5i uses vector relational data modeling (VRDM) to provide the Marine Corps with systems suitable for dynamic employment. MDOC5i uses the Global Information Network Architecture (GINA) as its VRDM platform. This research uses GINA to create a ubiquitous decision model that can be configured to USMC scenarios. The research implements the ubiquitous model and proves functionality through a network analysis use case. This decision model will serve as the base model for all GINA implementations. The ability to quickly construct and adapt GINA models based on scenarios and integrate those models into a common framework will provide the Marine Corps with information overmatch against future adversaries., Captain, United States Marine Corps, Approved for public release. Distribution is unlimited.

Current multi-domain command and control computer networks require significant oversight to ensure acceptable levels of security. Firewalls are the proactive security management tool at the network’s edge to determine malicious and benign traffic classes. This work aims to develop machine learning algorithms through deep learning and semi-supervised clustering, to enable the profiling of potential threats through network traffic analysis within large-scale networks. This research accomplishes these objectives by analyzing enterprise network data at the packet level using deep learning to classify traffic patterns. In addition, this work examines the efficacy of several machine learning model types and multiple imbalanced data handling techniques. This work also incorporates packet streams for identifying and classifying user behaviors. Tests of the packet classification models demonstrated that deep learning is sensitive to malicious traffic but underperforms in identifying allowed traffic compared to traditional algorithms. However, imbalanced data handling techniques provide performance benefits to some deep learning models. Conversely, semi-supervised clustering accurately identified and classified multiple user behaviors. These models provide an automated tool to learn and predict future traffic patterns. Applying these techniques within large-scale networks detect abnormalities faster and gives network operators greater awareness of user traffic., Outstanding Thesis, Captain, United States Marine Corps, Approved for public release. Distribution is unlimited.

In recent years, anonymization services such as Tor have become a popular resource for terrorist organizations and violent extremist groups. These adversaries use Tor to access the Dark Web to distribute video media as a way to recruit, train, and incite violence and acts of terrorism worldwide. This research strives to address this issue by examining and analyzing the use and development of video fingerprinting attacks using deep learning models. These high-performing deep learning models are called Deep Fingerprinting, which is used to predict video patterns with high accuracy in a closed-world setting. We pose ourselves as the adversary by passively observing raw network traffic as a user downloads a short video from YouTube. Based on traffic patterns, we can deduce what video the user was streaming with higher accuracy than previously obtained. In addition, our results include identifying the genre of the video. Our results suggest that an adversary may predict the video a user downloads over Tor with up to 83% accuracy, even when the user applies additional defenses to protect online privacy. By comparing different Deep Fingerprinting models with one another, we can better understand which models perform better from both the attacker and user’s perspective., Lieutenant, United States Navy, Approved for public release. Distribution is unlimited.

The Marine Corps’ 38th Commandant calls for a force that can “affect the behavior of real-world pacing threats” through its “innovation, ingenuity, and willingness to continually adapt to and initiate changes in the operating environment.” Having a healthy and inclusive environment is vital to the Marine Corps’ success in developing and maintaining a force that can rise to the higher standard. Currently, the closest snapshot of the health of Marine Corps culture is collected via command climate surveys, which contain quantitative and free-response questions. Senior levels of Marine Corps leadership do not gain value from the free-text data because 1) analysis on free-text comments is time consuming and no analysis is done on those comments for the commander, and 2) there is no trend analysis on the aggregate of surveys over time and throughout commands, so they are not commander specific. This thesis developed a tool to gain value from free-text comments; it applied natural language processing (NLP) techniques, specifically topic extraction and sentiment analysis, on free-text comments from command climate surveys to provide a quantitative analysis on the health of culture in Marine Corps units. This approach can also be replicated to other services and can inspire efforts that will positively impact talent management and retention., Captain, United States Marine Corps, Approved for public release. Distribution is unlimited.

In an increasingly digital world, online anonymity and privacy is a paramount issue for internet users. Tools like The Onion Router (Tor) offer users anonymous internet browsing. Recently, however, Tor's anonymity has been compromised through fingerprinting, where machine learning models are used to analyze Tor traffic and predict user viewing habits, with some models achieving an accuracy of over 99%. There are defenses for Tor that attempt to prevent fingerprinting, but many are defeated by new techniques that utilize Deep Neural Networks (DNNs). New defenses that are robust against DNNs use adversarial examples to fool the classifier, but those defenses either assume the user has access to the full traffic trace beforehand or require expensive maintenance from Tor servers. In this thesis, we propose Prism, a defense against fingerprinting attacks that uses adversarial examples to fool classifiers in real time. We describe a novel method of adversarial example generation that enables adversarial example creation as input is learned over time. Prism injects these adversarial examples into the Tor traffic stream to prevent DNNs from accurately predicting sites that a user is viewing, even if the DNN is hardened by adversarial training. We show that Prism reduces the accuracy of defended fingerprinting models from over 98% to 0%. We also show that Prism can be implemented entirely on the server side, increasing deployability for users who run Tor on devices without GPUs., Outstanding Thesis, Ensign, United States Navy, Approved for public release. Distribution is unlimited.

As web users resort to adopting encrypted networks like Tor to protect their anonymity online, adversaries find new ways to collect their private information. Since videos over the internet are a major source of recruitment, training, incitement to commit acts of terrorism, and more, this project envisions developing a machine learning algorithm that can help the Department of Defense find terrorists who take advantage of the dark web to help promote extremist ideology. This thesis describes the steps for training a machine learning classifier in a closed-world scenario to predict YouTube video patterns over an encrypted network like Tor. Our results suggest an adversary may predict the video that a user downloads over Tor with up to 92% accuracy, or may predict the length of a video with error as low as 5.3s. Similar to known website fingerprinting attacks, we show that Tor is susceptible to video fingerprinting, suggesting that Tor does not provide the level of anonymity as previously thought.

In recent years, anonymization services such as Tor have become a popular resource for terrorist organizations and violent extremist groups. These adversaries use Tor to access the Dark Web to distribute video media as a way to recruit, train, and incite violence and acts of terrorism worldwide. This research strives to address this issue by examining and analyzing the use and development of video fingerprinting attacks using deep learning models. These high-performing deep learning models are called Deep Fingerprinting, which is used to predict video patterns with high accuracy in a closed-world setting. We pose ourselves as the adversary by passively observing raw network traffic as a user downloads a short video from YouTube. Based on traffic patterns, we can deduce what video the user was streaming with higher accuracy than previously obtained. In addition, our results include identifying the genre of the video. Our results suggest that an adversary may predict the video a user downloads over Tor with up to 83% accuracy, even when the user applies additional defenses to protect online privacy. By comparing different Deep Fingerprinting models with one another, we can better understand which models perform better from both the attacker and user’s perspective.

Background and Purpose: To identify factors associated with prior stroke at presentation in patients with cryptogenic stroke (CS) and patent foramen ovale (PFO). Methods: We studied cross-sectional data from the International PFO Consortium Study (NCT00859885). Patients with first-ever stroke and those with prior stroke at baseline were analyzed for an association with PFO-related (right-to-left shunt at rest, atrial septal aneurysm, deep venous thrombosis, pulmonary embolism, and Valsalva maneuver) and PFO-unrelated factors (age, gender, BMI, hypertension, diabetes mellitus, hypercholesterolemia, smoking, migraine, coronary artery disease, aortic plaque). A multivariable analysis was used to adjust effect estimation for confounding, e.g., owing to the age-dependent definition of study groups in this cross-sectional study design. Results: We identified 635 patients with first-ever and 53 patients with prior stroke. Age, BMI, hypertension, diabetes mellitus, hypercholesterolemia, coronary artery disease, and right-to-left shunt (RLS) at rest were significantly associated with prior stroke. Using a pre-specified multivariable logistic regression model, age (Odds Ratio 1.06), BMI (OR 1.06), hypercholesterolemia (OR 1.90) and RLS at rest (OR 1.88) were strongly associated with prior stroke.Based on these factors, we developed a nomogram to illustrate the strength of the relation of individual factors to prior stroke. Conclusion: In patients with CS and PFO, the likelihood of prior stroke is associated with both, PFO-related and PFO-unrelated factors.

The desire to be more sensitive to the body, and to the embodied and embedded relationships that are constantly built in Academic life, brought us - a supervisor and a doctoral student – to share a common path of research, whose scope is to interrogate the role of the body in the academic experience. We started from the experience of breathing, a material and symbolic process. Where the body acts (we explored our practices), in fact, the ‘quality’ of breathing always becomes crucial. Inspiration/expiration, with 2 pauses in the middle. This quaternary structure is a metaphor of knowledge itself. Our breathing tells a story, on different levels, about the relationships between internal and external worlds. And this story is the basis of our research. Our research method is duoethnography (Sawyer, Norris, 2013). It incorporates theories of dialogue, life history, aesthetic representation, and transformation. Human experience is the ‘site of research’ (Norris, Sawyer & Lund, 2012), apt to foster an intimate experience, of reciprocal unveiling. This encourages collaboration and critical friendship, and challenges established roles. By juxtaposing our stories we achieved a provocative interrogation of our ideas and presuppositions about many aspects of our academic engagement. The experience itself of writing on our breathing co-created the sensation of ‘being alive’ in a traditional, formal, institutional, academic setting. Reflecting about our diversity reconnected us with our sense of humanity and with the deep mystery of eco-systemic life (Bateson). It opened new reflections about our daily interactions. In the end, we will ask to ourselves if and how a “Duoethnography of Breathing” really allows a different experience of university life, more embodied and embedded. A process of transformation (Mezirow), not only individual, which can re-enchant the relationship between student and supervisor in the academic setting