Smarter Supervision with the Help of AI Models and Autodetection

Article | 04-10-2022 | Web editor

CAN  wants to increase its focus on their data-driven approach to supervision, meaning  we will be guided more and more by data in our selection and detection processes. This is one of the goals in our Multi-Year Strategic Plan. The Port of Rotterdam almost holds  50 kilometres of containers on some days. Of course we cannot open and inspect  all of them. So we need to develop smarter techniques. We need  technology to  help us, based on data available to us, to select the most suspicious containers for inspection.

From simple to complex algorithms

An important concept in data-driven working:  Algorithms. An algorithm is a set of instructions  to achieve a certain goal, a kind of ‘recipe’ to solve a problem. The execution of the steps in an algorithm can be automated. That way an algorithm describes the steps that software must take to achieve a specific goal.

Already customs is using  algorithms. So far these are simple algorithms, based on if-then-else combinations. An example is  a ‘sanctions algorithm’:  if the information we have about a shipment shows that its goods are on a sanctions list, then we select this shipment for a more extensive inspection, and we stop the shipment. Else  we do not stop the shipment. Increasingly we want to use more complex algorithms, based on artificial intelligence (AI) models. We are currently developing AI models that can automatically recognise objects in X-ray images. We collect these images by scanning postal packages, we make  X-ray photos of the goods.

An AI model must first be trained to recognise certain objects. If  you want a model to recognise pills inside postal packages, the AI software needs to be ‘fed’ with as many, preferably hundreds, of X-ray images of postal packages containing different types of pills in all kinds of situations. The software is trained to distinguish those from postal packages containing no pills.

Learning to recognise with the help of artificial intelligence

At some point, the software starts to recognise the objects you have trained it for. This is artificial intelligence (AI), where an IT system mimics human intelligence to perform tasks and can improve itself during that process. The model from the above example ‘learns’ under the control of humans, who then check in test situations whether the model is sufficiently reliable. Our data scientists train a model as long as needed until that is achieved. This process is carried out with a lot of diligence, using enough data and the right data. The training rules are strict. Developing and training of AI models follows strict protocols.

Recognising objects in X-ray images is a form of ‘autodetection’, the automated collection and analysis of data to detect risks. We want to increase automated retrieval  of data and use these data for automated detection. In addition to X-ray scans, we want to use e.g. pictures from truck number plates, container GPS data, container radiation level data and more. These data can ‘cleverly’ be combined with declaration data in various ways to get even better results.

AI models can also be used on ‘structured’ data (texts). We can feed an AI model with data  from declarations, freight documents, certificates, company files, and data from previous shipments of the same product, such as shipping routes. With such data we can then train an AI model to detect ‘deviations’. The model can assess a shipment as ‘normal’: everything is going as expected. But also as ‘anomaly’. For example: for avocados on a cargo ship departing from location A, the route via location B to Rotterdam is the most logical one. This is the normal pattern. If a ship however takes a different route, the model recognises this as a deviation. For this kind of analysis, we need to cleverly combine all kinds of information. The more reliable data, and the more those data are related to the goal, the better the results will be.

Customs work remains a human’s job

There is also a different side to algorithms. Recently, they have been in the news for a less positive reason, especially in the context of work performed by government organisations. Algorithms are around in our daily lives already. They are used by Google, by dating sites, insurance companies and others. A form of positive use is to have algorithms helping human’s to make better decisions. This is true in our case: we deploy algorithms to support our decision making. The rules for using algorithms are very strict. As an additional safety measure, ultimately customs officers make the decisions, such as whether or not to stop goods. This way the human remains in control.

These  new methods do not replace the practised eye of a customs officer. A model cannot tell if  suspicious characters are loitering near a container. Or that a traveller at Schiphol Airport seems nervous. On the other hand: customs staff cannot manually go through all goods declarations. That is where AI models can make a big contribution. Human intuition remains important, but is increasingly supported by automated signals. The combination of both allows us to select the right shipments for a more comprehensive inspection. This increases the chances of intercepting for example dangerous toys, harmful counterfeit medicines or items that are on a sanctions list.

Improved supervision through AI and auto-detection

Customs has only just started developing AI models and applying these to autodetection. We are still training and testing models. So it may take a while before we can fully put these technologies into practice. They will allow us to continuously improve our supervision. AI models are becoming increasingly reliable when it comes to indicating the places where we are most likely to find something suspicious. We may not open every container and every parcel, but we have insights in the content of each of them.