Customs and Border Protection: IoT Gets The Goods
Artificial intelligence has played an important role in the massive improvements made in border security by many high-GDP countries over the past two decades. The border protection industry is now on the cusp of seeing major new products and technologies appearing that could result in more secure and efficient borders for both passengers and trade.
Customs and border protection management have undergone profound changes over the past two decades. Following the September 11 terrorist attacks in 2001, the United States and other countries scrambled to update their border management processes. The focus is on advanced screening to detect possible threats, as well as facilitating trade and travel as the volumes of both increased. With the need to screen passengers in advance of boarding an aircraft or scan a ship’s cargo ahead of arriving in a port, borders have increasingly become “virtual,” relying on data, intelligence, and risk management principles to secure them against threats.
Those in the industry have seen a shift from relying on manual checks, training, and instincts to-ward greater automation, application of artificial intelligence, and data analysis. Thirty years ago, Chris Thibedeau was working as a customs inspector for the Canada Border Services Agency (CBSA) in Nova Scotia – today he is the chief executive officer of TTEK, a customs and border protection management technology company. In those days, when a large vessel laden with thousands of containers arrived, analysts would be handed a large stack of paper bills of lading to pore over, looking for signs of high-risk shipments. It might be something as simple as a first time or unknown importer, cargo coming from countries that were a known source of illicit goods, or a shipment suspiciously directed to a PO Box – or it might be something that didn’ make sense, like automotive parts being transported inside a refrigerated container. Shipments meeting multiple “anomalies” were often deemed higher risk and worth investigation.
If you're just doing random selection, you never find anything.
Chris Thibedeau, CEO, Unknown Company
Through this process the stack of paper representing 1,000 commercial shipments might be whittled down to 100 individual shipments for closer scrutiny and inspection. These would be cross-referenced against other databases, including police and customs watchlists, looking for obvious red flags, such as an importer with a history of non-compliance or other enforcement actions. Human intelligence and open-source intelligence would also feed into the decision-making to determine which containers would be selected for inspection on arrival. Finally, the analysts would arrive at a shortlist of maybe just ten containers deemed to be highest risk. “Because we could only look at ten containers [in a single shift], we became very incentivized to make sure that any container we were going to offload at the warehouse that day was going to be the right one. If you were just doing a random selection, you were never going to find anything,” explains Thibedeau.
When Canada, the United States, and other nations moved to enhance border security post-9/11, many countries began to develop targeting systems that embraced the use of artificial intelligence principles to automatically rank shipments for inspection, drawing upon the same risk indicator rules and risk profiles developed by the old-school analysts.
TTEK is building several targeting centers for government agencies worldwide: Its library of over 65,000 proprietary rules targets shipments, using deductive and inductive logic as well as predictive, with automated risk-scoring driven by machine learning.
Thibedeau, who was then working at CBSA headquarters and was recruited to lead the development of its commercial risk assessment system called Titan, describes the jump from manually combing through stacks of paper to automated sorting of data as “a game changer” for border control agencies. “With the targeting system we could auto-mate deductive and inductive logic rules to identify known risk indicators and flag shipments deemed to be of highest risk to customs inspectors. We were putting the needles on top of the haystack,” he says.
Our company algorithms can sometimes outperform the human eye.
Marc-Olivier Roché, Smith Detection
While some nations, especially higher-GDP ones, use sophisticated targeting systems and the principles of risk management, many others still rely on less efficient processes. This usually results in extensive inspections of shipments, which can create bottlenecks in the ports of entry and ultimately result in delays and increased costs for importers. Thibedeau’s company TTEK is currently building several targeting centers for government agencies worldwide. It claims to have a library of over 65,000 proprietary rules for targeting shipments, using deductive and inductive logic as well as predictive, with automated risk-scoring driven by machine learning. He’s optimistic that the global market will quickly adopt this approach as benefits become clear. “We believe we’re on the leading edge of risk-based analytics,” he says.
Where There’s smoke
Artificial intelligence can also in-crease the effectiveness of border security equipment. Smiths Detection, a manufacturer of threat detection and screening technologies, began using AI to detect smuggled cigarettes inside cargo containers and trucks in 2008. Smiths uses convolutional neural networks to analyze images, produced by its scanners in ports, air-ports, and other sites, to improve detection of contraband, such as weapons, or to detect irregularities with trade. In some cases, the company’s algorithms can outperform the human eye, says Marc-Olivier Roché, Smith Detection’s product director for high-energy scanners: “We’ve been amazed at how the AI can detect weapons in a picture when an operator might look at the same picture and not see anything suspicious”.
Peekaboo, I see you: AI can detect weapons in a picture when an operator might look at the same picture and not see anything suspicious.
Increasing trade volumes and pressure on agencies to reduce or maintain staffing budgets were among the main factors pushing the development of image analysis automation, he says – but it hasn’t all been plain sailing. Early versions of the algorithms did not perform especially well, says Roché.
One challenge for many AI companies in the security space was obtaining images to use for training data because the industry-standard images can be hard to source or limited in number. How-ever, persistence has paid off for Smiths and additional training data quickly improved performance: for its weapon detection algorithm, for example, ingesting an additional 5,000 images reduced the false alarm rate from 13 percent to two percent, says Roché. Smiths is currently evaluating the use of AI to create artificial images to speed up deployment.
Crowd Control: Aurora AI has installed a systemin serveral European airports that allow travellers to mix in a single terminal and still maintain complete control.
Producing a workable solution re-lies on fine-tuning the output. An overly sensitive system will generate too many false alarms, while if not sensitive enough it risks letting contraband slip through. Typically, the best results come from customizing the algorithm’s output to an individual customer’s profile, says Roché. “Customers with more of a security focus will want a high probability of detection and care less about false alarms. Those with high throughput can’t afford having too many false alarms,” he explains.
There are still obstacles to adoption. In an industry with precise standards for a scanner’s performance, there are no standards to prevent manufacturers making exaggerated claims about their AI products, says Roché. Often, AI isn’t specified in tenders, while some clients worry that human operators may become lazy and rely entirely on the machine’s judgement. Finally, even if their system is responsible for a major bust – say of ivory at a port in Africa – the manufacturers typically can’t publicize results because agencies fear it would tip off smugglers to operational details.
Roché is confident new clients will still continue to be won over, especially as the technology becomes more effective and easier to de-ploy. The company is also working on additional uses, including being able to detect variations between the contents of a container and its description on the manifest. Initial work has been promising despite the complexity of cargo images. “Imagine a pallet of bananas sitting behind a pallet of oranges. It’s not straightforward for an algorithm to make sense of that,” he says.
AI makes facial recognition more suitable for the level of security in border control.
Nick Whitehead, Aurora AI
Putting names to Faces
The use of biometrics for travelers also opens up AI opportunities. Aurora-AI, a UK-based AI specialist in the aviation sector, has installed its facial recognition system in several major European airports, including at Manchester Airport and in two terminals at London’s Heathrow Airport. The system allows inter-national and domestic travelers to mix in a single terminal, offering significant efficiencies for airport operators. Having one set of shops and services and less infrastructure to keep passengers separated is attractive – but shared-occupancy terminals carry risks. A principal concern is that a passenger arriving on an international flight could directly board a domes-tic one to bypass immigration and customs, explains Nick Whitehead, Aurora-AI’s executive head of business development. To combat this, the company’s system uses cameras with infrared (IR) flash to photo-graph passengers as their boarding passes are scanned and they cross into the secure airside area.
We will soon be able to incorporate machine learning into all the passenger and cargo vetting systems.
Alan Bersin, Altana Trade
This AI process – verification – is able to be done with a much higher confidence level than identification, when a photograph of a person is matched against a database of individuals, and makes facial recognition using verification more suit-able for the level of security needed for border control, says Whitehead. Their system must basically guarantee it won’t allow unauthorized passengers to board whilst not falsely rejecting passengers, which creates delays. He confirms that he’s “highly confident” that his system is re-liable, and its efficacy is regularly tested by security staff.
Aurora-AI produced its first facial recognition algorithm around 2007, and Whitehead says it can identify an individual from a set of images as large as a single planeload of around 600 people to a standard suitable for border security. “As soon as you start trying to perform identification against a much larger group – say everyone coming to the airport in one day – then your chances of making a mistake go much higher,” he says.
The company also offers a wider suite of AI tools for the aviation sector, including using predictive analytics to generate forecasts. For example, its algorithm can use the schedule of arriving aircraft to predict the flow of passengers into the immigration hall, generating an ac-curate forecast of how many desks need to be open to maintain a certain queue time. “Using AI based on historic information can provide you with a better prediction than standard modeling,” says Whitehead.
Dirty Work at the Crossroads
Border security for passengers presents a huge data challenge as large amounts of data-sharing between partner nations and across national agencies “makes data veracity, lineage, and provenance tricky,” according to Mike Gormley, head of public sector at data unification company Tamr. Raw data collected by border agents may be inaccurate, misclassified, or other-wise “dirty” – for example, border agents may be in a rush as they enter in data, while travelers may struggle with language barriers when questioned or may intention-ally falsify information.
With targeting centers able to access multiple data sets – including advance passenger information submitted when booking an airline ticket, historical travel records, flight manifest data, and information on individuals in databases of known undesirables – it’s important to be able to match a passenger with the information on them held across those databases Since 2016, Tamr has been working on improving the entity resolution capabilities of the Global Traveler Assessment System (GTAS), a passenger screening system produced by the USA’s Customs and Border Protection (CBP), the country’s primary border control organization and the largest agency within the Department of Homeland Security (DHS). GTAS is freely available on GitHub for any country to use.
Perfect match: With targeting centers able to access multiple data sets – including advance passenger information submitted when booking an airline ticket, historical travel records, flight manifest data, and information on individuals in databases of known undesirables – it’s important to be able to match a passenger with the information on them held across those databases.
Tamr’s solution to improve entity resolution uses a variety of tools to clean data, including trigrams to work around typos, and techniques to improve matching of a passenger’s data between different data-bases to ensure passengers aren’t incorrectly matched (for example, if their name is the same or similar to one on a no-fly list), to build a clearer picture for border agents. Thorough cleansing and collation of detail gives border agencies an enhanced understanding of who is trying to enter their country, allowing faster detection and identification of known threats. It should also mean less unnecessary extra screening at the terminal and potentially shorter queues through the improved information sharing. “By including more data sets in an evaluation, agents can go multiple levels of information deeper to identify potential threats,” says Gormley.
Recent improvements in techniques means that the use of AI in border security globally is on the threshold of further large improvements in the next few years, says Alan Bersin, a senior fellow at Harvard’s Belfer Center and executive chairman of Altana Trade, an AI platform for improving international commerce. “I think we’re on the cusp of real breakthroughs in terms of big data analytics and machine learning,” he says.
Bersin, a former US CBP com-missioner, thinks one of the big problems with targeting will be overcome by federated learning. Agencies and companies in the private sector have always been reluctant to share their data, because of either privacy or proprietary concerns. Federated learning addresses this by allowing machine learning algorithms to access individual data sets without having to commingle them with others or download them to a central store. Keeping each learning source discrete negates the privacy issues that have made organizations unwilling to share data in the past. “That is a revolution in data processing,” Bersin says.
Currently, the CBP’s National Targeting Center operates on a rules-based system, but increasingly AI’s predictive powers will be used to generate new rules, processing massive data sets in real time and adding an additional layer to the existing targeting intelligence. “Eventually there will be a transition to a full AI solution and we will be able to incorporate machine learning into all of the targeting functions and also the passenger and cargo vetting system,” says Bersin, noting that “It’s a work literally in progress now.
“He believes this approach will eventually be adopted by many countries. “I think you’re going to see these AI, fully automated [targeting] systems increasingly show up in border management over the next two to five years,” he predicts.
Sign up for the Avnet Silica Newsletter!
Stay up-to-date with latest news on products, training opportunities and more!
Take a DEEP look into the future!
Get the latest market trends and in-depth trainings on our Digital Event Experience Portal!
Avnet Silica Design Hub
Browse and review hundreds of proven reference designs to accelerate your design process. Our designs can be modified and saved in our AVAIL design tool and then exported to your CAD tool of choice.