United States

Information and influence operations (IO) are a tool for advancing adversarial agendas in near-peer conflicts by augmenting narratives that drive actions in communities. As evidenced by the 2016 US Presidential Election, nation-state actors employ a variety of tactics, techniques, and procedures (TTPs), such as bot nets (semi-automated) and troll farms (human-powered), to conduct IO via text-driven communications. Countering IO has become a priority for the national security community. However, the national security community currently lacks a capability that can emulate the large-scale, adversarial IO activity, limiting our ability to train appropriate responses. Here we present a case study in using GPT-2, a state-of-the-art Transformer-based generative language model, as a synthetic opposition force for real-time IO activity emulation. We show that fine-tuning GPT-2 models with specific datasets can effectively emulate different personas in an information landscape. We differentiate model personas by measuring various language dimensions – emotion, sentiment, morality – using VaderSentiment and custom BERT-based classifiers. We further evaluate the generation quality using measures of local and global coherence, sentence quality, and generation velocity. Our initial results demonstrate that generative language models will be successful in emulating various personas engaged in a coordinated IO campaign. Further, these results suggest research should be conducted to understand appropriate mechanisms for prompting (i.e., feeding communications to a generative language model to spur additional content generation) these models in a way that furthers the desired narrative. We anticipate that the effective application of fine-tuning and prompting techniques will support the use of generative language models as a synthetic opposition in large-scale training exercises. Operating these models as a synthetic adversary will spur advances in the human-AI collaboration and training fields that improve our understanding of command-and-control (i.e., white cell to red force) and action-reaction (i.e., blue force to red force) relationships in IO response training. These advances will be useful in extending this synthetic content generation capability to other domains and media.

NAML 2021

Realistic Cacophony: Creating Synthetic Personas with Diverse Sentiment, Emotion, and Morality for Influence Operations Training

influence operations training

synthetic personas

generative language models

Naval Information Warfare Center Pacific is partnering with AFCEA San Diego chapter to host the workshop.

The NAML workshop is an opportunity for researchers from Navy research and development centers and our partner organizations to exchange technical ideas and to come together as a team.

The goals of NAML are:

* To spread awareness of current machine learning research relevant to Naval applications

* To connect machine learning researchers with experts in military requirements

* To build and strengthen collaborations within the defense research community

**Watch this video to hear about the AFCEA purpose.**

<iframe width=1280 height=706 frameborder="0" scrolling="no" src="https://screencast-o-matic.com/player/creleVVcvRh?v=6&ff=1&title=0&controls=1" allowfullscreen="true"></iframe>

poster

Accurate positioning and navigation are critical to a wide variety of Navy and DoD applications. A Global Positioning System (GPS) and an Inertial Navigation System (INS) make up the backbone of modern navigation systems. Even the best inertial sensors will drift over time without some sort of correction. GPS typically provides this correction, but is not always available due to environmental or adversarial factors. When this is the case, many systems and missions are limited by the inertial system’s holdover time (the length of time the system is able to provide useful position solutions without corrections). The DIRT-I project aims to increase this holdover time by applying reinforcement learning techniques to the inertial system’s position solution.

DIRT-I: Drift Improvement through Reinforement Learning - Inertial Navgation Systems

Autonomous systems have an array of specialized sensors to observe their surroundings. An autonomous platform must be able to combine many disparate streams of information in order to leverage all of the available information while creating a comprehensive model of its environment. Deep learning-based object detectors typically need to be trained on swaths of labeled data to achieve good performance. Unfortunately, collecting and labeling multimodal data in the real world is exceedingly labor-intensive. The use of video game graphics engines in the production of images and video has emerged as a relatively cheap and effective way to create new datasets. A unique aspect of using gaming engines to generate data is the ability to introduce domain randomization which randomizes certain parameters of the game engine and generation scheme in order to improve generalization to real-life data. In this paper, we outline the creation of a multi-modal synthetic dataset using domain randomization. Our dataset will focus on the two most popular sensors in autonomous vehicles, LiDAR and RGB cameras. We will discuss the effect of different parameters on each modality and perform baseline testing of an object detector using a data-fusion deep learning architecture on both our synthetic dataset and the KITTI dataset for comparison.

Multi-modal Dataset Generation using Domain Randomization for Object
Detection

We present initial research on physics-mathematics modeling, computational simulation, and data visualization that aim at discovering the physics of the adversarial minds of threat actors who develop exploits and malware. In this research, the physics of an adversarial mind refer to laws that govern the dynamics of human cognition as reflected in malware. The physics of adversarial minds, in turn, may enable us to discover cognitive vulnerabilities that are exploitable through defensive cyber deception to persistently detect malware across many generations, or for as long as the malware are developed by threat actors whose cognition we have analyzed and have found to be vulnerable. We aim at discovering physics of cognition that pinpoint innate and hence possibly involuntary cognitive vulnerabilities, as well as developmental cognitive vulnerabilities. We expect these cognitive vulnerabilities to be persistent and recurring in a variety of manifestations. The enabler of the discovery of cognitive vulnerabilities in cyber space is a deep and thorough understanding of how the cognition of threat actors and cyber terrain interact with each other to produce exploits and malware. We aim at discovering the interplay between computing and the cognitive processes that take place in the mind of threat actors. In cognitive science, human cognition is often characterized via mathematical models. Our goal is to extend and adapt these models to operations in cyber space, and thus integrate them with elements of computing. A fusion of cognitive phenomena and computing in a mathematical form serves as basis for this research, which is envisioned to leverage cyber mechanisms to learn and then probe the cognition of threat actors for cognitive vulnerabilities. A key approach in this research is the small-scale decomposition of the presence of human mind in malware into particles. We develop small-scale mathematical models that characterize the individual action of those particles, their physical quantities, and how those physical quantities are related to each other. We encode in these small-scale models the inner workings of cognition in ways that resemble physical matter such as gasses and fluids, while representing the physics of cognition with high fidelity. We then exploit small-scale models to develop large-scale mathematical models of malware-reflected cognition. We discover physical properties of an adversarial mind in aggregate from a probabilistic examination via statistical physics and nonlinear dynamical systems theory of the underlying logical particles. Of particular interest are macroscopic physical quantities of cognition that emerge from the collective behavior of particles. In other words, this research discovers the thermodynamics of an adversarial mind using malware as measurable observable. This research explores statistical physics and dynamical systems theory similarly to how these disciplines were used to study gasses and fluids in conventional physics. By developing high-fidelity mathematical models that establish equivalence between cognition in malware and physical matter, this research creates potential for inheriting the past success of those methods.

Mathematical Discovery of the Physics of Adversarial Minds in Malware for Customized and Targeted Deception

In a cyber defense scenario, it is important to mitigate threats on a system
in an automated and timely manner. While previous work with single-agent
reinforcement learning attempts to accomplish these goals, we improve upon
scalability by introducing a cooperative multi-agent reinforcement learning al-
gorithm. More specically, we implement a Deep Q-Network (DQN) algorithm
with Metropolis consensus weights based on shared action-value function pa-
rameters. Using a cyber simulation, we compare the results of our multi-agent
DQN to the results of a single-agent DQN.

Cooperative Multi-Agent Reinforcement Learning for Cyber Defense

Labeled real world data is often difficult to obtain and especially scarce in Navy-related domains. Currently, relevant annotated data that does exist is frequently limited to large or medium sized bounding boxes, making it difficult to train computer vision algorithms to recognize smaller objects of interest. In this work, we present a naval-specific video dataset of helicopter operations performed at sea. This dataset contains videos from multiple camera sensors to incorporate variations in lens distortions and camera noise. It consists of videos ranging from one to three minutes each recorded during Littoral Combat Ship (LCS) exercises off the California coast in the fall and winter. Special consideration was taken to emphasize small instances of helicopters relative to the field of view and therefore provides a more even ratio of small-, medium-, and large-sized bounding boxes for training more robust detectors and trackers. Following the conventions of the field, we define small, medium, and large objects as objects with bounding boxes sized: less than 32, 32-96, and greater than 96 pixels squared respectively. We benchmark these videos on object detection with special consideration given to small-object mean average precision.


Helicopter Video Dataset for Detecting and Tracking

Fielding autonomous robotic systems in safety- and mission-critical applications requires careful scrutiny and oversight. Deploying these systems is predicated upon the availability of technologies – such as Human Intent Prediction (HIP) and Artificial General Intelligence (AGI) – that currently do not exist. We prescribe a new approach to HIP in order to enable Safe Human-Robot Interactions (Safe-HRI). We examine the limitations of existing technologies, while recommending caution in the design, development, and deployment of such systems in safety- and mission-critical applications.

Towards Intent Prediction for Safe Human-Robot Interaction

Electronic Warfare Support (ES) is the act of detecting, characterizing, and classifying potentially hostile electromagnetic signals, a crucial component of self-protection for manned and autonomous platforms in the increasingly congested modern battlefield. Current ES methods perform early identification of radio waveforms emitted by comparing the incoming waveform to heavily curated libraries. While effective, this legacy approach is slow to respond to a novel and agile waveforms emerging from recent advances in software defined radios (SDRs). The Neural Arrays for Autonomous Electronic Support (NAAES) program, sponsored by ONR, is using machine learning to dramatically reduce the time and effort to respond to novel waveforms. We have demonstrated the ability to characterize and classify RF waveforms in real time using the Intel Neural Compute Stick 2 (NCS2). NAAES achieved a DNN-enhanced modulation classification rate of over 7500 pulses per second, using seven 1-2W NCS2 devices running a combined 750 GFLOP/s.

Neural Arrays for Autonomous Electronic Support

The discussion provides a general overivew of the DON analytical approach, the Jupiter Enterprise Data Environment, the DON Data Community of Practice, and the 2021 DataCOP Challenge.

Jupiter & the DON Data Community of Practice

This presentation is an overview of the recently developed SpikePropamine meta-learning framework submitted for publication in the Frontiers In Neurorobotics journal. Material from the original paper as well as more recent developments will be presented. 

Our goal in developing this framework is to leverage synaptic plasticity in Spiking Neural Networks (SNNs), the learning mechanism found in biological neural networks. Today, most engineered neural network only use fixed-weights despite the superior learning capabilities found in biological plastic neural networks. Therefore, SpikePropamine is a framework that harnesses the power of offline learning of fixed weights along with the online continual learning through synaptic plasticity. Specifically, along with the fixed weights, SpikePropamine additionally learns the rules that govern the online plastic learning and neuromodulation in SNNs through offline gradient descent. The capabilities of the framework are demonstrated through a series of challenging benchmarks. The parameters for variants of BCM, Oja’s and other plasticity and neuromodulated rules are learned. Results show that the SpikePropamine based networks are able to solve temporal learning tasks that fixed-weight SNNs fail to solve. Additional results highlight the robustness of SpikePropamine based SNNs to noise and novel conditions not seen during training. This is seen through minimal degradation of performance in a high-dimensional legged locomotion task governed by a SpikePropamine based SNN.


SpikePropamine: Meta-Learning Synaptic Plasticity for Spiking Neural Networks

It is well known that Deep Neural Network (DNN) implementations take massive volumes of data to sufficiently train their complex network architectures to perform high accuracy inferencing. The Machine Learning (ML) communities, academic and industry, have created sophisticated models of DNNs trained on millions of data samples to achieve robust classification performance. In the area of image analysis, to address the need for data, numerous efforts have led to standard curated datasets, such as MINST, CIFAR-10, Open Images, and ImageNet. In fact, the ImageNet dataset has become a de-facto image dataset for new algorithm evaluation. In a yearly competition, participants are invited to submit new and novel DNN architectures for assessment. The architectures are trained using the ImageNet database and then evaluated for classification performance. In order of development, some of the well-known winning networks originating from this competition are AlexNet, ZFNet, GoogleNet, VGG Net, ResNet, and Inception V4. As a benefit to the deep learning community, many of the competition submissions are released as open-source for others to use and refine. It is important to recognize that the key to the performance of these networks is that the data input into the network is similar in nature to the data on which the network was trained. Herein is the problem facing many Department of Defense (DoD) ML implementations; the lack of sufficient, operationally relevant data to train and innovate DNN architectures for the DoD community. 
As part of a 2020 Navy Small Business Technology Transfer (STTR) program, Areté and its academic partner, the University of Arizona, are developing an interactive scenario building tool capable of generating realistic synthetic 360° videos in real-time for use in DoD training simulators. This novel capability is accomplished by combining the latest advances in generative adversarial networks (GANs), Areté's expertise in physics-based modeling, and the advanced simulation capabilities of the Unreal Engine. In this talk, Areté will present examples of maritime relevant synthetic data compared to real-world data. Areté will then demonstrate how the data can be used to train DNN architectures directly or through transfer learning and how key technologies are marshaled together.


Realistic Cacophony: Creating Synthetic Personas with Diverse Sentiment, Emotion, and Morality for Influence Operations Training

Next from NAML 2021

DIRT-I: Drift Improvement through Reinforement Learning - Inertial Navgation Systems

Similar lecture

Artificial Intelligence Applications to Cybersecurity (AI ATAC) Prize Challenges I&II

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES