The latest ransomware attacks, and their many victims, have led to feverish efforts to find ways of providing protection. The results achieved with Artificial Intelligence have gained a lot of attention. More and more security products are being promoted on the basis of this technology, but the term covers some very different approaches.
It normally happens only in films: computers battle each other, try to get past each other’s security features and gain control over their opponents. Last year, seven networked systems competed to try to hack into each other and take out the $2 million prize for the winner of the Cyber Grand Challenge.
The computers taking part in the challenge analysed their opponents’ network traffic, independently wrote entirely new programs, repaired security loopholes and modified attack tools. Artificial Intelligence, or AI, was the magic word that gave the machines the ability to learn what they needed, and enabled them to follow paths that hadn’t been pre-programmed for them. The excitement following the competition was huge, since the systems discovered security loopholes that were entirely unknown up to that point.
From expert knowledge to machine learning
If Artificial Intelligence had not been a top focus of IT security before this event, it certainly has been since. But it isn’t a new technology by a long shot: the heyday of Artificial Intelligence was back in the 1980s, when it involved expert systems that were designed to reflect fuzzy knowledge using extensive rule sets. Such knowledge cannot be formalised; it is often generated based on experiences and usually covers a number of fields of knowledge.
One example is medical diagnoses, in which experienced and inexperienced doctors will often produce very different results. If the doctors have to incorporate the patients’ personal circumstances or mental states, they need to draw on additional knowledge from other professional fields. The extremely complex structure of the “knowledge base” has proved to be a disadvantage in this regard. A handful of experts need to spend a quite lengthy period of time documenting their knowledge in the form of rules that computers can work with – which can take weeks or even months.
This led to the idea of developing algorithms capable of building up a knowledge base on their own, in other words, capable of learning. Neural networks are among the most successful machine learning processes. They have also been around since the 1980s, but have only become suitable for everyday use thanks to the power of modern processors. Neural networks are the attempt to represent in digital form the knowledge of how nerve cells function that has been obtained from research into the brain. Essentially, turning zeroes and ones into neurons and synapses.
Neural networks generate their knowledge through training with sample data. Only once they have been fed with countless examples can they learn what’s important, and only then can they make appropriate decisions. To do this, the network gets feedback on its performance with each new training example, which it then uses to keep on adjusting its various parameters until the functionality for particular tasks has been established. That could mean recognising images or speech, or identifying malware or breaches.
Previously, customers of providers of traditional antivirus products had to rely on laboratories to discover and investigate bugs. Only at that point could traditional agents be prepared at the end points to recognise them. That means that traditional solutions can act only against known attacks and malware. The superiority of neural networks, on the other hand, lies in the fact that they can recognise new variations with no need for them to have been analysed in the laboratories run by the security specialists.
In the network and on the host
AI products are available in various forms. Agents on terminals are just as common as special network components, known as appliances. While the former try to detect harmful files and unusual operations in the operating system, the latter examine data traffic in the network for anomalies. But not every procedure is suitable for everything. Ransomware is easier to detect on the host than in the network, but if a harmful program that has penetrated the network scans the LAN for interesting devices or open ports, a box listening in from within the network will spot it more effectively than an agent on the host.
The form that training takes also differs: some neural networks are regularly trained with new bugs at the manufacturer’s end, until their switching is arranged in such a way that they will recognise these attack patterns. Unlike traditional products, however, the neural network is then capable of identifying an entire category of new malware of a given type, not just an individual instance. The manufacturers buy in extra malware for this type of training.
Duelling neural networks
Other products train themselves using data packets discovered in the company network. The behaviour patterns occurring in these networks will often depend on the company. Machine learning also takes place quite unobtrusively in the Internet: major network operators and Internet service providers run global platforms to recognise anomalies in global Internet traffic, to make it possible to respond rapidly to attacks with a global ambit.
Thanks to the success of machine learning, almost all manufacturers of security software now highlight the use of AI methods to promote their software products, and it has become difficult for customers to distinguish between them. Even so, development continues: the latest thing in AI research is “adversarial networks”, in which two networks face up to each other, the second evaluating the results from the first, and effectively mirroring the battle between the systems in the Cyber Grand Challenge.
You will also find news about all aspects of it-sa and the IT security environment in the it-sa Security Newsletter.