United States

NAML 2021

Robustness and Safety Q&amp;A Session

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>

Robustness and Safety Q&A Session

live discussion

This work describes the problems that can occur when Autonomous Systems have not been designed with security in mind. Examples are given of vulnerabilities found in Autonomous Systems - the risks that these systems face, both from malicious and naive users are described. Baseline security measures and advice are given for designing, and using Autonomous Systems in a way which minimises the risk of a vulnerability being exploited. This security advice does not cover “artificial Intelligence” attacks (such as adversarial inputs, or data poisoning), these attacks form another research topic entirely. Instead, we focus on basic software hygiene, as relevant to Autonomous Systems. Since these systems are typically network connected, this falls into the area of Cyber security hygiene, following best practice to enhance security. Cases of incorrect configurations, poor security choices and risky assumptions are all demonstrated. 
© Crown copyright (2021), Dstl.

How good Cyber Hygiene can help defend Autonomous Systems

Traditional machine learning pipelines require data to be centralized. For data that are collected and stored on multiple edge devices, or for data owned by multiple institutions that must not share their sensitive information, the traditional pipeline becomes impractical due to bandwidth bottlenecks and privacy concerns. Federated learning is a machine learning specialization that addresses these issues and allows for accurate, timely decision making on data that are not centrally available. Rather than aggregating distributed data into one centralized dataset, federated learning assumes that the data remain at the edge. By doing so, federated learning reduces the bandwidth strain over communication lines, allows institutions to share knowledge without giving up their data, and leverages distributed edge computing. In our current studies, we seek to develop a federated algorithm that converges quickly to an accurate model, that is practical for implementation on edge devices with limited processing capabilities, and that is bandwidth efficient. More specifically, we focus on federated optimization with one central server and many clients, and introduce our algorithm called Federated Kronecker-factored quasi-Newton. We are using Kronecker-factorization to make loss-curvature information available cheaply at the edge for federated training in this second-order optimization procedure, which to our knowledge, has not yet been done. We highlight the benefits and challenges of second order optimization in the federated setting and demonstrate preliminary performance curves that show improved model accuracy on smaller amounts of information exchanged and in less training time than some state-of-the-art algorithms. We also show that even though our algorithm is a second-order optimization technique, it is still practical to implement.

Second-Order Federated Optimization

The increasing capabilities and decreasing costs of cameras has greatly expanded the amount of image and video data collected by analysts. Often, there is far more data than manual analysis can handle, and training personnel with the required expertise can be costly and error-prone. Computer vision and machine learning can help, but creating an operationally viable solution usually requires deep learning expertise, large amounts of labeled training data, and software engineering skills.

To address these challenges, we have developed an open-source toolkit, Video and Imagery Analytics for Marine Environments (VIAME; github.com/VIAME). VIAME includes a wide range of image and video analysis capabilities such as target detection and classification, tracking, image classification and image registration. VIAME enables the rapid integration of new visual analytics, and includes user interfaces, databases and evaluation/scoring capabilities. More significantly, VIAME includes the capability for users to create new analytics, specific to their problems, through user interfaces, without any programming or knowledge of deep learning. Through video search and interactive query refinement, users can quickly build a model for a novel classification problem by providing positive/negative feedback on search examples suggested by the system, and then run it on any amount of imagery or video. For more challenging problems, users can manually annotate images and then train a deep learning detection and classification capability.

While originally created for maritime organism analytics (and funded by NOAA National Marine Fisheries Service), on ongoing ONR and AFRL grants we have recently added a number of features to its interfaces specific to aerial video processing including multiple vessel detection and tracking, high-value target tracking, 3d-model generation around vessels, and a new web interface for training detection or tracker models to compliment existing desktop interfaces.


VIAME - An Open Source Video Analytics Toolkit for the Maritime Environment

Detecting Cyber Invasions with DNS Analytics & DGA Detection Leveraging Machine Learning

AFCEA Communities of Interest

"Each year at the Naval Postgraduate School (NPS), more than 1,600 active duty servicemembers attend a wide variety of resident and distance learning (DL) graduate and doctoral programs.  This student body represents some the most innovative and passionate minds currently serving the Department of Defense (DoD). Nearly all of these students must complete a capstone project or thesis as a requirement for graduation. However, typically only a small portion of the student body has insight into their services’ current capability development challenges or pressing research needs. While the student body struggles to find relative work that furthers the military’s research priorities, senior leaders across the DoD have no way of discovering the depth, breadth, and diversity of knowledge, experience, and skills represented by the officers attending graduate education nor seeing the research or innovation occurring at NPS and other military academic institutions like it.  To address this challenge, we are proposing a cloud-based, collaborative research platform that enables the discovery of completed, on-going, and future research opportunities for researchers, research sponsors, and leaders to connect, collaborate, and deliver timely, prioritized information to DoD leadership in order to support informed decision-making.  

The Vision: 
This platform, that we call Athena, is being built to leverage artificial intelligence techniques and machine learning to maximize both user experience and taxonomy development. The functional test version that our team has built includes: 
Primary Features:   
- Discover research databases that contain completed, on-going, and future/potential research topics 
- Connect users and organize Communities of Interest (COI) via an “idea marketplace”  
- Collaborate across the DoD and with Industry Partners on research and technologies  
- Deliver viable products that informs leadership decision-making and research prioritization"

Athena - The cloud-based Collaborative Research Tool

MIT Lincoln Labs Disaster Response Image Analysis

AWS Data Labs Disaster Response Imagery Training

Unmanned Undersea Systems (UUXs) often operate in communication-limited environments, which require some degree of autonomy in order to operate efficiently and effectively. Current autonomous systems often operate as “black boxes”, which may offer high performance but lack transparency. As such, the use of autonomous systems requires a high degree of trust from the user, which may be difficult to gain. Explainable artificial intelligence (XAI) may offer a way to build user trust in AI systems, by increasing transparency and offering explanations for anomalous behavior. Currently, XAI is used by researchers as an analytic tool to identify poor model performance and improve user interfacing. In this work, we propose that explainability can be transformed into a user-facing tool to foster understanding in both users and developers. We will demonstrate this concept by contextualizing a system’s actions within a restrictive environment, utilizing a simplified scenario with simulated sensor inputs. Success and failure will be defined by creating metrics for the succinctness and relevance of an explanation, including both the salience of provided context and the absence of important information. The
resulting explanations will piece together the environmental conditions, mission, behaviors, and operational status in order to deliver an intuitive explanation of the decisions of an autonomous system.

Exploring Trust and Explainability of Autonomous Systems in an Operational Environment

Event cameras utilize novel imaging sensors patterned after visual pathways in the animal brain that are responsive to low contrast, transient events. The pixels of Dynamic Vision Sensors react independently and asynchronously to changes in brightness, creating a stream of time-stamped events encoding the pixel location and sign of the brightness change.  Furthermore, they have extremely high dynamic range, low latency, and lower power requirements, making them ideal for persistent surveillance in varying light conditions on platforms where size, weight, and power (SWaP) are at a premium.  

Their output is fundamentally different from traditional frame-based cameras, where each pixel is sampled at the same time at a fixed rate.  As such, traditional methods of machine learning for computer vision cannot be directly applied.  We have been developing a maritime data set using these cameras for the purpose of automated ship detection, and employing methods based on optical flow for object detection and tracking. In this talk, we describe our data set and initial results.

Premium content

Robustness and Safety Q&A Session

Downloads

Next from NAML 2021

How good Cyber Hygiene can help defend Autonomous Systems

Similar lecture

Resilient Distributed Adaptive Cyber-Defense Using Blockchain