We hear about it every day but what is artificial intelligence really? Jordi Iparraguirre, EURid Innovation Manager , shared with us his views on the limits and vulnerabilities related to artificial neural networks to also show the flip side of technological evolution.
Artificial intelligence is now part of our daily life, we all talk about it. But what is artificial intelligence really?
Artificial intelligence (AI) is the ability of a machine to display human abilities such as reasoning, learning, planning and creativity. For example image and text recognition, grouping of elements into classes, speech recognition, identification of complex patterns based on thousands of parameters, etc.
The programming model used in artificial intelligence systems is different from that used in traditional systems. In traditional systems, you receive an input, you write a set of sequential rules to process that input, and you get an output. For example, to calculate the VAT for a given product, you enter the price of the product as an input, multiply that value by the tax to be applied (%), and you get the amount of tax (€) to be applied. to that specific product. By entering a different price, we will obtain the tax to be applied to a different product.
Artificial intelligence techniques, on the other hand, work differently. Initially, the system is “trained” by entering both the input data and the desired output. For example, the system is shown images and the system is taught that these images show apples, other images show oranges and so on. The artificial intelligence system is then asked to develop, on its own, a set of rules that allow it to correctly classify a new image. However, if the machine is shown a type of fruit different from that shown to the system initially, or a fruit already inserted in the machine but in a context or with characteristics different from those shown during the “training” phase of the machine (for example an apple blue, instead of green),
Artificial intelligence techniques have several branches of study and development. A few years ago they were called “expert systems” and used a specific set of algorithms and languages. Today, neural networks and machine learning have taken over using new algorithms that are possible thanks to the ever increasing computing power and huge amounts of data offered by a digitized society.
We always talk about the advantages related to artificial neural networks, what are the main limitations and how are they addressed?
Neural networks are a powerful tool for handling unstructured data, such as images or sounds. Their theoretical aspects are not new. The first computational model was developed in 1943 with further developments starting in the seventies with the first computers in research centers. But the speed and memory limits of the computers of the time severely hampered their use and adoption. It was only at the beginning of the new millennium (even if it would be better to speak of a “century”, to be less dramatic) that processors and memory devices acquired the capacity to handle the enormous amounts of data necessary to offer practical solutions to the beyond theoretical studies.
Today, neural networks are the basis for the functioning of applications and solutions that we use every day and that are even found in our mobile phones. However, despite being able to perform their function, it is often difficult to understand the reason that explains the reason for a certain result.
We are able to understand the mathematics that governs the whole process, we begin to understand the mechanisms of action of the different levels of artificial neurons, but it is not yet clear, for example, what is the number of levels actually needed to make the result optimal. . In some cases we proceed by trial and error, following a path that is more reminiscent of art than engineering. Despite this, commercial applications of neural networks move millions of euros and there are groups of experts working on methods that will help us understand them and obtain an explainable artificial intelligence , where it is possible to know the reason for a certain result.
We are witnessing a growing use of AI and intelligent systems applied to different areas. How do attackers exploit vulnerabilities and weaknesses? Are there any relevant examples to share?
There are several techniques called “Adversarial”.
If you know the type and quality of the data used to train the artificial intelligence system, it is possible to develop countermeasures to “confuse” the system.
An interesting example can be that of t-shirts with images and colors that can fool an artificial intelligence system for identifying people. Another example is that of stickers pasted on road signs, used to fool car cameras that read road signs. There are also systems to sabotage spam filters and thus get unwanted messages to our inbox.
An important aspect to consider in artificial intelligence is that of the so-called “prejudices” or “bias” (it may happen that some systems are discriminating, such as those, for example, that are not able to identify people with darker skin) . Many times this is due to improper design and training of the artificial intelligence system. Artificial intelligence systems need to be re-educated and be able to accept feedback from trusted users or sources to adapt to changes in the environment and the data they process.
After all, the most important aspect is represented by the quality of the data used to train the artificial intelligence system. For example, it may happen that the number of examples inserted in the machine is not sufficient, that examples are inserted whose interpretation is easily subject to distortion, or that the machine has been instructed with poor quality data. A phrase often used in the field of computer science is “Garbage In, Garbage Out” (GIGO), that is, a result of poor value (garbage out) derives from a set of incoming data of poor quality (garbage in).
How is artificial intelligence applied in EURid?
EURid uses systems based on artificial intelligence at various levels. One, and perhaps the best known, is the award-winning APEWS (Abuse Prevention and Early Warning System) which analyzes the likelihood of a domain name being used for illegal purposes. In this case, the domain name is in any case registered, but not delegated. Our legal department initiates a procedure to verify the identity of the owner. If the owner does not respond or is unable to prove their identity, the domain is suspended.
Another area in which EURid applies artificial intelligence is the automatic classification of websites associated with .eu domain names. For example, knowing how the owners of .eu domain names use them allows us to better develop future marketing campaigns.
