���˴�Ƭ

ISR is a broad term that covers the collection and processing of data to produce intelligence to inform decision-makers. The data from ISR systems can take many forms including: 

• Visible, radio-frequency (RF) and infrared (IR) imagery
• RF and IR signatures
• Structured and non-structured intelligence information.

ISR systems produce vast quantities of data. The reliance on human-operator/analyst input is a major bottleneck in many Defence and security applications. Furthermore, the volume, diversity and complexity of ISR data now exceed a human operator's ability to recognise more complex patterns; particularly transient and non-repetitive patterns. 

Machine learning is very effective at recognising patterns, deriving correlations and identifying anomalies. By applying machine learning to the large volumes of IRS data it is possible to extract important details more easily, reducing operator overload and improving situational awareness. 

AIML has worked with , Snap Surveillance and to develop machine learning techniques to:

• Detect, classify, track and in some cases identify targets from ISR data
• Automatically transform imagery into geospatial information 
• Enhance the accuracy of target recognition and tracking in crowded, complex, resolution-limited and noise-limited data 
• Search through large databases of data quickly 
• Perform behaviour modelling and anomaly detection in large-scale surveillance networks
• Implement semantic change detection to identify significant changes in a scene from a set of images
• Perform content-based image searches
• Perform entry- and relationship-extraction from unstructured data; for example, intel reports, running sheets, intercept data, social media and other open source information
• Recognize patterns and derive correlations between data from various sources.

Electronic Warfare involves detecting and jamming an opponent’s electromagnetic sensing or communications capabilities (over optical and radio-frequencies). 

Historically, radar systems have been built around specifically “tuned” emitters with well-defined, fixed characteristics and a limited range of operational modes and waveforms. The trend toward digital, programmable radio-frequency (RF) equipment (epitomised by software-defined radio) means that radar characteristics can be changed dynamically, creating unique signatures on-the-fly or even mimicking commercial RF signals. Furthermore, the proliferation of low cost, size, weight and power RF devices means that the RF environment is becoming increasingly congested. This means that hostile emitters are becoming harder to locate and identify; let alone jam and confuse. The integration of machine learning and artificial intelligence into radar target detection, acquisition, and tracking algorithms has the potential to greatly enhance the ability of a radar to detect and respond to anomalous detections - aka cognitive EW.

AIML is working with to apply machine learning techniques to:

• Automatic modulation classification techniques to classify and identify various modern radar waveforms
• Target classification in the presence of anomalous interference

A further complication in the modern EW battlespace is that, with the proliferation of passive radar, Active Electronically Scanned Arrays (AESA) and Low-Probability-of-Intercept (LPI) radar systems, it is becoming increasingly difficult to detect and geolocate radar emissions. This eliminates a vital source of situational awareness for Defence personnel, placing additional emphasis on other techniques to not only provide situational awareness but also aid in increasing situational understanding. Specifically, multi-sensor data fusion at the platform-level and network-enabled information fusion over the entire battlespace.  Machine learning algorithms are very effective at correlating data from multiple sources, recognising patterns and identifying of anomalies. Machine learning will form the core of the systems that are built to provide improved situational understanding in the future EW battlespace.

Machine learning forms the foundation of autonomous systems, including:

• Autonomous vehicles (including land, sea and air vehicles)
• Automated digital assistants
• Tactical Decision Aids (TDA).

AIML has an established track-record for applying machine learning techniques to the navigation of autonomous vehicles. AIML has developed machine learning techniques for: 

• Simultaneous Localisation and Mapping (SLAM)
• Robust long-term autonomous navigation

AIML’s experience in visual question answering, natural language processing and deep learning has direct applications to automated digital assistants that can be used to: 

• Aggregate, collate and interpolate both formatted and unformatted data from a wide variety of sources
• Recognise patterns and anomalies 

An important war-fighting paradigm is decision-superiority; the ability to make better decisions faster than an adversary. Traditionally, human operators, analysts and decision-makers have performed the necessary tasks to understand a situation and determine an appropriate course of action. As the volume of data increases TDAs are needed to extract important information from data; to help the humans make-sense of the information and to provide suggestions of the best course of action. Machine learning techniques and algorithms will form the core of these TDAs. 

A key issue in the development of AI-powered TDAs is the ability of a human operator to “trust” the interpretation of the data and the course of action it suggests. For a human to trust the TDA’s conclusions and recommendations, the TDA must be able to explain how it has derived its knowledge from the information it has collected, transformed and fused: this is often referred to as explainable AI, which is a subset of the broader concept of trusted autonomous systems. At AIML the development of trusted autonomous systems is a key research priority. 
 

Cybersecurity is the protection of internet-connected systems, including hardware, software and data, from cyber-attacks. 

Machine learning has been successfully applied to security monitoring activities. With machine learning, it is possible to parse large quantities of data such as network traffic and log files in near-real time, to undertake “pattern-of-life analysis” and look for anomalies. Leveraging this ability, machine learning can be used to detect cyber-attacks in progress, characterise the attack traffic and, with appropriate training, automatically counter that attack. The ability of machine learning to check for anomalous behaviours lends itself to use in chasing ransomware. Machine learning can also be used to perform structural code analysis: the process of analysing code looking for common security loopholes and vulnerabilities.