Artificial Intelligence in Security

In the late 1970s, academics and industry has worked on the subject of AI and with the advent of numerous connected networks and big data, Neural Networks and Deep Learning have emerged as key application areas of AI based. Last 24 months have been interesting for the Security Industry with new age technologies such as Artificial Intelligence, Machine Learning and Machine Vision making mainstream appearances; products getting smarter & sophisticated; and the overall adoption roadmap of these technologies seems much wide spread now than ever.

The growing number of companies innovating in AI applications for physical security is a clear indication that these are more than a passing phase.

These technologies have been identified as ‘Exponential Technologies’ as technology for which performance relative to cost and size is doubling every year.

The concept of Artificial intelligence is not new at all. It has been at the core of our smartphones since the last five years in the form of Siri, Alexa and Assistant (for Apple Amazon and Google environments respectively). For that matter, let’s go back a decade: ‘Text to Speech’ is a classic example of consumer genesis of AI.

This ‘Natural Language Processing’ will go beyond just speaking into the phone and resulting in typed text, rather these smartphones and other similar devices will even understand what has been entered. With smart elements such as smart tagging, key word identifying and geographical tagging: these systems will correlate the inputs and provide a better topic identification. A sentence such as ‘someone forced entered through main door’ will not be read just the way it sounds according to the English language, but these machines will apply the technology to decipher the outcome / output result by identifying it to be an ‘act of intrusion’.

A major deep learning breakthrough occurred in 2015, which radically reduced the machine vision error rate. During a machine-vision competition, a large visual database designed for use in visual object recognition software research - succeeded for the first time in surpassing the five percent average human error rate, when analyzing a database of images. Such a rapid enhancement was caused not only by progress in advanced algorithms, but also by the development of new and much faster hardware systems based on graphics processing unit (GPU) cores, instead of traditional central processing units (CPUs). These new marriage of hardware and software architectures allowed for faster learning phases.

Understanding these Technologies

Artificial intelligence (AI) is the body of science, algorithms, and machines that are able to perform some version of learning and independent problem solving, based on advanced software and hardware components. Data is core to AI; large datasets are the foundation for recent performance improvements across market applications. Within the AI subject field there are other sub-fields such as machine learning, neural networks, and deep learning.

Machine Learning involves collecting large amounts of data related to a problem, training a computer using this data and employing this model to process new data. Deep learning, which is a branch of Machine learning, is a way to emulate the functions of the human brain using software algorithms.

With deep learning, you can show a computer many different images and it will “learn” to distinguish the differences. This is the “training” phase. After the neural network learns about the data, it can then use “inference” to interpret new data based on what it has learned. For example, if it has seen enough cars before, the system will know when a new image is a car.

For this to happen, the system “learns” by looking through large volumes of data that need faster processing time to achieve artificial intelligence (AI). This is where Graphic Processing Unit (GPU) comes in, and is making artificial intelligence accessible to the security industry. By improving surveillance efficiency and accuracy with AI during the course of time - the technology is already being added to existing cameras to more efficiently search for objects or persons of interest. In surveillance applications, AI could eliminate the need for humans to do hard laborious work such as look at hours of video footage.

A system relying on neural networks differs from conventional pattern-recognition systems, in that it will continuously learn from experience, and base its ability to discern and recognize its surroundings like human beings do: by learning from real sounds, images, and other sensory input.

Enabler of Competitive Advantage

In the recent past, the usage of machine learning – primarily in the applications for image processing and natural language processing, machine translation has exploded in the industry and allied sectors. Most of these capabilities are based on open-source libraries, and can be deployed easily and rather cost effectively in the cloud or distributed cloud – the entry barriers are on an all-time low and is on the rampant decline further. This will always give way for product and services innovations and the benefits will be beyond standard automation based tasks.

The proliferation of these technologies have especially resonated with the c-level business leaders who have embraced these in order to differentiate themselves from competitors.

As per a recent study conducted – 45% of businesses have deployed these technologies in the financial sector that are core to their business and about 10% are still experimenting – thus taking the overall count to more than 50% of the total businesses who are leveraging these technologies at some level.

Let’s take financial and human capital organisations for example – the adoption of AI and ML lies in the motivation to saving money – where in these technologies cut costs and increase productivity by a factor of billions of dollars. However, the main focus of these technologies over a period of time is well timed and informed decision making. About 60% organisations use it to drive decision making – and data is at the core of all of this. Any innovation that makes better use of data, and enables data scientists to combine disparate sources of data in a meaningful visualisation that could have the potential to gain competitive advantage. The areas where these are used are: Risk Management, Performance Analysis & Reporting, Ideation and Automation.

The key adoption drivers are to these are: Extract better (read ‘meaningful’) information, increase the pace of productivity, reduce overall operating costs and extract more value from data available (or data mined).

AI and the Security Industry

AI, ML and Deep learning technologies offer a much higher level of accuracy and reliability to recognise an object or behaviour, and accurately classify significantly higher than the traditional / conventional rules-based algorithms.

New age algorithms make deep learning solutions to view a scene as intuitively as a human being. The ability of deep learning algorithms to view a scene intuitively, as a human viewer would, means that detection accuracy increases dramatically. Neural networks allow a computer to apply a series of algorithms to a given situation learning to identify increasingly more sophisticated features such as shapes, colors, tones, textures - unlike rules-based solutions which are limited to the limited inputs. Also, as compute power continues to increase, the neural networks will leverage for improved accuracy.

Deep learning has demonstrated its capacity to increase the effectiveness of a computer to reliably classify objects and behavior. Security software companies are now marketing analytics that can leverage deep learning to turn vast amounts of video footage into usable information in a fraction of the time it would have taken in the past. In the video surveillance analytics market, some algorithm developers are using AI in the form of deep learning to maximize their output efficiency. Most of these software companies are training the algorithm mostly in the cloud, using solutions such as Amazon Web Services or Azure as heavy computing power is expensive so it is leased. This has bolstered the growth of the security could market by a factor of 35% yearly.

Programming and coding methods have also been optimised for rapid deployment. In the past, the speed and quality of the analytic was directly impacted by the team’s size. Now, analytics can be developed for niche applications quickly and with far lesser resource requirement boosting up competition amongst niche players. Furthermore, most of the large companies are focused on one-size fits all, leaving opportunity in niche applications.

Video processing software also allows users to interact with the surveillance footage using a search engine like interface with natural language search terms such (such as ‘man in a red shirt’, or ‘blue car’). This makes searching video like a breeze and much easier to use and search, there by significant reducing time and human effort to pull footage from hundreds of thousands of cameras. In addition to this, the ability to detect multiple objects and classify them allows for much greater insight – for example, this extends the ability to recognizing a cars color, type, make, model, and analysing which direction and speed it is moving at - making it possible to provide trending analysis,  draw patterns and conclusions based on historical data.

This coupled with the concept of Command & Control Centres, the industry is moving in the direction of ‘Usable and Actionable’ Result oriented framework for security information management and security operations.

Face Recognition Applications and Biometrics:

Most facial recognition analytics on the market today feature some kind of deep learning. Not only does it increase the accuracy of facial recognition sensors, it also enables faces to be identified in larger and more crowded scenes. In the wake of recent terrorist attacks in crowded locations, this capability could change the whole approach to security monitoring, allowing law enforcement to track suspects with greater speed and efficiency.

However, deep learning analytics are doing more than just improving accuracy rates. They are also enabling the system to make assumptions and provide business intelligence on a detected face. Age and sex recognition algorithms, which are particularly popular within retail applications, allow end-users to profile potential customers and target marketing material appropriately. Furthermore, some vendors claim to be able to recognize a person’s emotions through analytic algorithms. There remains some debate as to the accuracy of these solutions currently.

One area where facial recognition has the potential to disrupt, is in the access control market. Facial recognition solutions have been used for a number of years at passport control in airports. However, as the price of the technology and cameras reduces, it is expected that facial recognition will be used to prevent access to restricted areas.

Risk Based Access Control:

In nearly all access control systems, the authentication process is a singular event; a credential is presented and access is granted or denied. However, by leveraging the advanced capabilities of neural networks this process can be made more intelligent.

At a basic level, risk based access control can be summarized as follows: in many sports stadiums public office space is combined with the actual sports stadium. Access to the office locations needs to be open when workers are accessing it during the week. However, when the sports event is on this access should be restricted and only provided to key personnel with the correct access control credentials.

In the example given the authentication process is dynamic, depending on circumstances. This allows the system to shift gears and provide a higher level of building security when certain events happen. By introducing the intelligence of AI to a risk-based access control authentication decision, the process can be made more complicated. Instead of defining risk levels by looking at the events currently on in the building, the system could pull data from other security or building management systems or social media alerts to make decisions based on this data. The main barrier to developing this level of complexity using traditional rules-based analytics is that there are simply too many variables to account for. The use of neural networks means developers do not need to write rules for the system to follow, they simply need to provide the algorithm with objectives and training data. Over time, the system will be able to decide how all of the inputs should relate to the current risk level.

Real Time Crime Analysis

Predictive analysis tools have come a long way since the first products and solutions. Police agencies can now make use of a wide range of data inputs and advanced data mining techniques to predict where criminal activity is likely to occur. This approach is called the predictive crime center and it is becoming an important part of modern policing.

Data from video surveillance, traffic management cameras, audio analytics, gunshot detection, weather systems, and other public safety systems are analyzed in parallel to identify patterns and potential threat events. Over the past few years, social media has also become a viable tool in public safety. In many cases, incidents are first reported on social media platforms such as Twitter and Facebook. Analyzing the data “hose” is challenging but can support quicker emergency response times when applied successfully.

The future predictive command center will be reliant on powerful analytics software to extract intelligence from the unstructured data sources routed into command centers. The solution must be open to ensure that enough data is fed into the big data solution. Artificial Intelligence will play a key role in navigating this data, recognizing patterns and making intelligent decisions independently of the human operators.

Related, a more localized solution to predictive alerts could be implemented in the suburban environment. AI could be used to identify criminal behavior such as burglary and theft. Deep learning video surveillance cameras could support the process, alerting to triggers such as loitering, repeatedly walking past the same spot and wearing clothing that makes it hard to identify an individual. Artificial intelligence in this case could also make better use of the crime data available – in terms of frequency, time, approach, and direction, to more accurately predict future criminal events.

Managing Large Data Volumes (Big Data), Privacy and Cyber Security

AI supports more detailed statistical analysis of the operations of security departments. However, one of the challenges for “big data” is in having enough data to make reliable statistical conclusions. As we have already highlighted, the short-term data analysis opportunities will likely be in situations where large data sets are created such as in safe city projects. Smaller companies may not be able to make accurate decisions based on a more limited data set. Ultimately, AI is required to identify that thing that does not belong in the data: this requires enough data to recognize anomalies, not just new content.

Another challenge is in normalizing the data. Social media represents an important new source, but in order to alert to abnormal behavior there needs to be an assessment of what is normal behavior. Consequently, normalizing the data set is critical to assess what is the typical amount of conversation around a specific topic.

The physical security market could also learn from other industries in applying AI to customer service. Natural language processing is improving and chat bots are increasingly used by consumer facing companies to provide an artificial intelligence interface for their users. This type of technology could be deployed to support physical security and employee security applications in the future.

As discussed, the performance of an artificial intelligence algorithm is linked to the quality and size of the available dataset. In an increasingly connected world, new physical security sensors are being deployed all the time, driving different data types into the AI solution. While this is great for the evolution of deep learning algorithms, it does present a threat in terms of cybersecurity. Historically, the biggest challenge for cybersecurity in the physical security market has been the lack of awareness throughout the route-to-market. End-users often underestimated the force of cyber threats and integrators and equipment suppliers were not focused on building cyber protection into their solutions.

More recently, equipment vendors in the video surveillance and access control markets have shown more commitment towards cyber security. Responses have included product hardening guides, encryption certification, the auditing of firmware code and partnerships with dedicated cyber security solution providers.

However, many of the connected device start-ups entering the physical security market don’t have the resources to focus on cyber security and will remain a threat to the overall solution. As the large IT companies improve their cyber defenses, it is likely that attacks on IoT vendors will increase in regularity and intensity. Ultimately, these companies will provide an easier target to hackers looking to maximize their impact.

AI is also relevant in terms of cyber security technology. For example, the defense market is already using deep learning applications to analyze cyber data to better protect critical national infrastructure. Many industry observers think that hackers will use this technology to escalate their attacks in the future too, increasing the cyber threat further. There are also some concerns related to how AI solutions interpret inputs and whether this could be manipulated by cyber criminals to cause confusion and damage in the future.

Related to the cyber threat, data privacy and risk will also be important considerations as AI solutions become more pervasive. Data encryption of video surveillance images is not common at the moment; mostly it is used in healthcare or critical infrastructure. GDPR (General Data Protection Regulation) in the EU could define video as unique personal data which may change the encryption requirements for video feeds.

GDPR could also impact what data is stored and how it is shared, impacting the analytics market. In particular, face recognition and person classification analytics will be considered personal data and have constraints on what can be done with the information. In response, Belgium announced that it will be banning the use of facial recognition for private use. The legislation does allow access control law enforcement applications but is an example of how data privacy could impact the physical security AI market.

Future is NOW for Artificial Intelligence

AI will be disruptive to many industries – not just the physical security market. Moreover, the impact of AI will not be just on industries and finance, but our entire society, especially in the areas of privacy and data security, labor, and ethics. The need for data security and privacy is more essential today than ever, given the availability of such powerful technologies.

The physical security market is primed to benefit from AI for two reasons:

• AI, in the form of deep learning algorithms, has the potential to revolutionize the video surveillance analytics market providing face recognition, object recognition and behavior recognition at a reliability level that will really matter to end-users.

• The physical security industry generates data. Video surveillance images, access control data, audio analytics, social media, police records management systems and other IoT sensors all generate data that can be correlated and analyzed by artificial intelligence systems to build a safer society. Intelligently managing this data is huge challenge. AI can help solve this problem.

The challenge for physical security vendors, end-users, and integrators will be how to make the most of the AI opportunity. This will involve investment, education and judgement to best apply this transformative technology to the individual challenges faced by each participant.