United States

MITRE ATLAS is a knowledge base of adversary tactics, techniques, and case studies for machine learning systems based on real-world observations, demonstrations from machine learning red teams and security groups, and the state of the possible from academic research.

ATLAS enables researchers to navigate the landscape of threats to machine learning systems and is modeled after the MITRE ATT&amp;CK® framework. The tactics and techniques in ATLAS are complementary to those in ATT&amp;CK.

ATLAS can be used to:
- Develop case studies for your internal red teaming exercises
- Build defensive threat assessments from ATLAS TTPs
- Inform your security analysts of emerging threats to AI systems
- Report real-world adversary use of AI/ML techniques
- Understand adversary behaviors and mitigate threats to AI systems 

In this talk, we&#39;ll describe the ATLAS framework, give real-world examples of it&#39;s use, and give recommendations for based on our experience for future discovery of vulnerabilities in AI systems and related projects.

NAML 2022

ATLAS (Adversarial Threat Landscape for AI Systems)

vulnerabilities

adversarial ml

machine learning

security

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>

NAML 2021

MITRE ATLAS is a knowledge base of adversary tactics, techniques, and case studies for machine learning systems based on real-world observations, demonstrations from machine learning red teams and security groups, and the state of the possible from academic research.

ATLAS enables researchers to navigate the landscape of threats to machine learning systems and is modeled after the MITRE ATT&CK® framework. The tactics and techniques in ATLAS are complementary to those in ATT&CK.

ATLAS can be used to:
- Develop case studies for your internal red teaming exercises
- Build defensive threat assessments from ATLAS TTPs
- Inform your security analysts of emerging threats to AI systems
- Report real-world adversary use of AI/ML techniques
- Understand adversary behaviors and mitigate threats to AI systems 

In this talk, we'll describe the ATLAS framework, give real-world examples of it's use, and give recommendations for based on our experience for future discovery of vulnerabilities in AI systems and related projects.

technical paper

##### For full functionality of Underline, please SHUT OFF your VPN, or use a personal device.


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


<figure class="video_container">
  <iframe src="https://s3.amazonaws.com/pf-user-files-01/u-59356/uploads/2022-03-22/hk13mxo/NAML_ReceptionWithtitle.mp4" width=826  height=526  frameborder="0" allowfullscreen="true"> </iframe>
</figure>

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



This talk describes the principle of ``General Cyclical Training'' in machine learning, where training starts and ends with "easy training" and the "hard training" happens during the middle epochs. We propose several manifestations for training neural networks, including algorithmic examples (via hyper-parameters and loss functions), data-based examples, and model-based examples. Specifically, we introduce several novel techniques: cyclical weight decay, cyclical batch size, cyclical focal loss, cyclical softmax temperature, cyclical data augmentation, cyclical gradient clipping, and cyclical semi-supervised learning. In addition, we demonstrate that cyclical weight decay, cyclical softmax temperature, and cyclical gradient clipping (as three examples of this principle) are beneficial in the test accuracy performance of a trained model. Furthermore, we discuss model-based examples (such as pretraining and knowledge distillation) from the perspective of general cyclical training and recommend some changes to the typical training methodology. In summary, this paper defines the general cyclical training concept and discusses several specific ways in which this concept can be applied to training neural networks.


On General Cyclical Training of Neural Networks

Derivations are given of various versions of equilibrium propagation using adjoint methods. The learning rules are studied using simulations with the MNIST data set.

Variational Perspective on Learning Rules for Networks with an Energy Function

With the advent of quantum supremacy, quantum technologies are establishing themselves profitable with near-term commercial products becoming more and more prominent in different parts of the world, with applications that include the military sector. Among such applications, the area of quantum machine learning (QML) is starting to establish itself as a field of its own. Much like classical machine learning, the goal is to be able to generalize a process into multiple applications, but a deeper question remains as to whether there is computational advantage against classical processes via quantum-inspired algorithms; common ways to find such advantages involves proving that there is no best known classical algorithm that performs more efficiently than a quantum algorithm, and processing wise that there exists a run- time exponential advantage over classical computing . While proving quantum advantage and supremacy is controversial and non-trivial, an additional challenge lies in the processing limitation imposed by currently available quantum computers, often referred to as Noisy Intermediate Scale Quantum (NISQ) processors . Such limitations have forced developers to compile algorithms limited both in the num- ber of qubits and number of qubit operations by a hardware-specific metric referred to as Quantum Volume (QV). As a result, it is widely accepted in academia and industry that there is not clear quantum advantage over classical processors given the current technological readiness of available quantum processors due to noise and quantum bits qubit de-coherence limiting the number of qubits and gate operations. However, while acknowledging such limitations, there are some avenues to continue advancing the paradigm of quantum machine learning applications acknowledging hardware limitations: 1) theoretical explorations are done to demonstrate, via com- putational complexity theory, implementations that could get attained in the future once a fault-tolerant quantum computer becomes available, 2) via hybrid instances where some machine learning tasks are managed by a quantum computer while a classical processor takes the rest, such that, quantum machine learning algorithms are used to classify quantum states that are produced by a quantum simulator, and 3) programming a NISQ computer and assessing its quantum algorithm performance against a classical processor counterpart.

Quantum Assisted Machine Learning (QAML) - Promises and Challenges

In their strategic documents, the Commandant of the Marine Corps and the Chief of Naval Operations call for increased investment in artificial intelligence (AI) with the goal of achieving decision superiority. To responsibly deploy a deep learning model in a high stakes operating environment, we must know when these models are unsure or are providing information that is akin to an unsubstantiated guess. Deterministic deep learning models employed today provide no such indicator; Bayesian Deep Learning fills this gap and is a current AI research focal point. Bayesian deep learning models provide answers to classification and regression problems as well as associated uncertainty metrics. Moreover, these models perform on par with or outperform their deterministic counterparts. At the NPS Applied Artificial Intelligence Lab, research with Bayesian models is being conducted on several Navy-relevant problems using large scale datasets, achieving state of the art results. In this work, we show our results from three studies. In the first study, we focus on a single element passive sonar sensor and use more than a year of passive sonar data labeled through automated fusion with automatic identification system (AIS) information. Our Bayesian models classify multiple ships simultaneously (multi-label classification) with a weighted F1 score of 0.8 and provide the classification uncertainty. In the second study, we focus on using on-surface ship recordings to classify internal soundscapes including activation of onboard pumps. We demonstrate that a rise in the classification uncertainty correlates well with the transitions in the soundscape (transient sounds). In the third study, Bayesian models were used to predict temperature brightnesses from infrared images for ingestion by data assimilation models used by Fleet Numerical for improved weather forecasting. These models predict temperature brightness within 4 degrees on 95% of the test data. These results generalize to temporally independent datasets, an important indicator of potential performance in a live setting. Putting Bayesian deep learning models into Naval production-level systems would contribute to decision superiority in three ways. First, for problems where some amount of uncertainty is acceptable, such as sensor to sensor emulation, the level of uncertainty can be adjusted to control the quality of the emulation. Secondly, for high-risk problems, such as target classification, the uncertainty measures provided by deep Bayesian models can be used for real-time system failure detection (i.e., target misclassification). Lastly, these uncertainty measures allow for detection of a distribution shift due to sensor degradation or environmental changes, an indicator that a model needs to be retrained or that a different model is required. In all three cases, the problem of making a decision based on an unsubstantiated guess has been eliminated.

Bayesian Deep Learning for Navy Applications

We perform a study on the usefulness of cyclostationary features of digitally-modulated waveforms for signal classification. We conduct exploratory data analysis to examine the interactions between and the behavioral changes in higher order moments and cumulants across a range of waveform configurations. We apply ten supervised machine learning methods to measure model prediction accuracy on waveform modulation types based solely on cyclostationary features. In addition to evaluating the trade-off space of accuracy vs. computational efficiency of each classifier, we also explore four dimensionality reduction techniques for cluster analysis. Our empirical results suggest that cyclostationary features provide useful information for categorizing signals and clustering digitally-modulated waveforms. Lastly, we conclude with recommendations for future work.

How useful are Cyclostationary Features in Classifying Digitally-Modulated Waveforms

Modern radar simulation development primarily uses raw data collects from blue forces placed in intelligence databases that have not yet been thoroughly vetted and distilled into a usable data product.  These resources often have redundant, ambiguous, or missing information describing modes and operation of the system that must be preprocessed by those local to the organization developing the simulations rather than the Intelligence Communities (ICs) themselves.  The ICs’ process to generate a Tier I or II product takes considerable time and doesn’t always meet the timeline requirement to support the Warfighter in a dynamically changing Electronic Warfare (EW) threat environment.  Hence, it is required for personnel to utilize data-driven tools leveraging Artificial Intelligence (AI) and Machine Learning (ML) to down-select redundant or ambiguous information and impute missing information to construct quick-reaction models at a relatively high frequency.  This effort employs clustering algorithms to find redundant or ambiguous modes for a particular emitter and then constructs the radar simulation based on the consolidation of modes.  The quality is comparable to the traditional manual construction of simulations in many cases and drastically increases the speed and quantity at which simulations are processed.  Missing parameter data is addressed by utilizing a decision-tree based regressor that imputes the missing parameters based on training data taken from other sources that could provide inference (i.e. other radars).  Lowering the variance of these estimates requires execution on data at fairly low levels describing the radar system, but has potential to initially bridge parameter gaps not yet filled. 

Machine learning applications for quick-reaction development of radar simulations with imperfect data

As one of the pillars of readiness, maintenance manning plays a critical role in the production of mission capable aircraft; the quantity, skills, and experience of maintainers onboard directly impacts the ability of a squadron to meet its readiness goals. Current distribution of maintainers is performed manually according to a set of heuristics and expert knowledge. Prior work performed under the Digital Aviation Readiness Technology Engine (DARTE) effort focused on automating this process via a statistical approach to identifying targeted, data-driven personnel moves expected to increase the readiness of the enterprise. Improving upon this work, the Resource Optimization with Agent Reinforcement (ROAR) model is built using reinforcement learning to train an agent which directs maintainer transfers within the fleet. The agent learns within a simulated environment which incorporates predictive models of important features into the transition model. This approach allows the agent to consider additional elements of the state of the enterprise (e.g. expected change in mission capable aircraft). In this way, the agent can select targeted personnel transfers which maximize readiness across the enterprise.

Resource Optimization with Agent Reinforcement

Machine learning research generally necessitates extensive data requirements. In the sub-field of computer vision, there is a high demand for visual data that accurately depicts the real world around us. This data has traditionally taken the form of digital images captured with optical sensors (physical cameras) and annotations applied during post-processing. Synthetic imagery provides a mechanism to dynamically create this training content for machine learning algorithms.
This project’s goal was to bring advanced machine learning technologies to explosive ordnance disposal (EOD) teams. The application of synthetic imagery technology can assist in the detection, recognition, and identification of ordnance while increasing the level of safety and efficiency for EOD specialists in the field.

Synthetic Imagery Applications in Computer Vision

Current artificial intelligence (AI) systems excel in narrow scoped closed worlds such as board games and image classification. However, AI systems performance drops when accommodating to constantly changing conditions which are commonly faced by warfighters in the real world. To begin solving this limitation, open world novelty detection, characterization, and accommodation technology is needed to solve military relevant problems. We collaborate with DARPA SAIL-ON (Science of Artificial Intelligence and Learning for Open-world Novelty) performers using probabilistic programming, zero-shot plan updating, meta-model based adaptation, and knowledge graphs in multiple capstone projects to demonstrate the technology applicability.

DoD Applications of Novelty Accommodating Artificial Intelligence

We are applying a combined approach that utilizes generative adversarial networks (GANs) and transformers towards a general solution for data deconflation. Data deconflation expands on classical blind source separation problems, such as that of the Cocktail Party Problem, to superposition of data objects not confined to signals. Separating the conflated data objects (e.g aligning voices to their speakers, separating information from multiple co-located sensors, etc) into their individual sources allows for meaningful analysis of the data. Although classical solutions have proven successful in a certain problem set, it is unclear whether these solutions can be generalized. Our current approach works with a simple illustrative example for subsequence separation from a conflated sequence. The GAN’s objective is to be able to generate each subsequence with a label identifying its source, while being able to differentiate between properly and improperly generated sequences. We will present on our current status and the full network architecture and discuss application to more real-world examples. 

[1] Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., & Courville, A. (2017). Improved Training of Wasserstein GANs. arXiv [cs.LG]. Opgehaal van http://arxiv.org/abs/1704.00028 

[2] Hallman, R., & Cybenko, G. (2021). The Data Deconflation Problem: Moving from Classical to Emerging Solutions. In Proceedings of the 6th International Conference on Internet of Things, Big Data and Security. Special Session on Artificial Intelligence for Emerging IoT Systems: Open Challenges and Novel Perspectives. SCITEPRESS - Science and Technology Publications. https://doi.org/10.5220/0010530403750380 

[3] Sutskever, I., Vinyals, O., & Le, Q. V. (2014). Sequence to Sequence Learning with Neural Networks. arXiv [cs.CL]. Opgehaal van http://arxiv.org/abs/1409.3215

ATLAS (Adversarial Threat Landscape for AI Systems)

Downloads

Next from NAML 2022

On General Cyclical Training of Neural Networks

Similar lecture

Data Management for Artificial Intelligence and Machine Learning Implementation Across the Department of the Navy

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES