Artificial Intelligence (along with Machine Learning) seems to be the hottest buzzword in just about every segment of the IT industry nowadays, and not without reason. The very idea of teaching a machine to mimic the way humans think (but much, much quicker) without the need to develop millions of complex rules sounds amazing: instead, machine learning models are simply trained by feeding them with large amounts of carefully selected data.
There is however a subtle but crucial distinction between “thinking like a human” (which in academic circles is usually referred as “Strong AI” and to this day remains largely a philosophical concept) and “performing intellectual tasks like a human”, which is the gist of Artificial General Intelligence (AGI). The latter is an active research field with dozens of companies and academic institutions working on various practical applications of general AI. Much more prevalent, however, are the applications of Weak Artificial Intelligence or “Narrow AI”, which can only be trained to solve a single and rather narrow task – like language processing or image recognition.
Although the theoretical foundations of machine learning go back to the 1940s, only recently a massive surge in available computing power thanks to cloud services and specialized hardware has made it accessible to everyone. Thousands of startups are developing their AI-powered solutions for various problems. Some of those, like intelligent classification of photos or virtual voice assistants, are already an integral part of our daily lives; others, like driverless cars, are expected to become reality in a few years.
AIs are already beating humans at games and even in public debates – surely they will soon replace us in other important fields, like cybersecurity? Well, this is exactly where reality often fails to match customer expectations fueled by the intense hype wave that still surrounds AI and machine learning. Looking at various truly amazing AI applications developed by companies like Google, IBM or Tesla, some customers tend to believe that sooner or later AIs are going to replace humans completely, at least in some less creative jobs.
When it comes to cybersecurity, it’s hard to blame them, really… As companies go through the digital transformation, they are facing new challenges: growing complexity of their IT infrastructures, massive amounts of sensitive data spread across multiple clouds, and the increasing shortage of skilled people to deal with them. Even large businesses with strong security teams cannot keep up with the latest cybersecurity risks.
Having AI as potential replacement for overworked humans to ensure that threats and breaches are detected and mitigated in real time without any manual forensic analysis and decision-making – that would be awesome, wouldn’t it? Alas, people waiting for solutions like that need a reality check.
First, artificial intelligence, at least in its practical definition, was never intended to replace humans, but rather to augment their powers by automating the most tedious and boring parts of their jobs and leaving more time for creative and productive tasks. Upgrading to AI-powered tools from traditional “not-so-smart” software products may feel like switching from pen and paper to a computer, but both just provide humans with better, more convenient tools to do their job faster and with less effort.
Second, even leaving all potential ethical consequences aside, there are several technological challenges that need to be addressed specifically for the field of cybersecurity.
- Availability and quality of training data that are required for training cybersecurity-related ML models. This data almost always contains massive amounts of sensitive information – intellectual property, PII or otherwise strictly regulated data – which companies aren’t willing to share with security vendors.
- Formal verification and testing of machine learning models is a massive challenge of its own. Making sure that an AI-based cybersecurity product does not misbehave under real-world conditions (or indeed under adversarial examples specifically crafted to deceive ML models) is something that vendors are still figuring out, and in many cases, this is only possible through a collaboration with customers.
- While in many applications it’s perfectly fine to train a model once and then use it for years, the field of cybersecurity is constantly evolving, and threat models must be continuously updated, expanded and retrained on newly discovered threats.
Does it mean that AI cannot be used in cybersecurity? Not at all, and in fact, the market is already booming, with numerous AI/ML-powered cybersecurity solutions available right now – the solutions that aim to offer deeper, more holistic real-time visibility into the security posture of an organization across multiple IT environments; to provide intelligent assistance for human forensic analysts by making their job more productive; to help identify previously unknown threats. In other words, to augment but definitely not to replace humans!
Perhaps the most popular approach is applying Big Data Analytics methods to raw security data for detecting patterns or anomalies in network traffic flows, application activities or user behavior. This method has led to the creation of whole new market segments variously referred to as security intelligence platforms or next-generation SIEM. These tools manage to reduce the number of false positives and other noise generated by traditional SIEMs and provide a forensic analyst with a low number of context-enriched alerts ranked by risk scores and often accompanied by actionable mitigation recommendations.
Another class of AI solutions for cybersecurity is based around true cognitive technologies – such as language processing and semantic reasoning. Potential applications include generating structured threat intelligence from unstructured textual and multimedia data (ranging from academic research papers to criminal communications on the Dark Web), proactive protection against phishing attacks or, again, intelligent decision support for human experts. Alas, we are yet to see sufficiently mature products of this kind on the market.
It’s also worth noting that some vendors are already offering products bearing the “autonomous” label. However, customers should take such claims with a pinch of salt. Yes, products like the Oracle Autonomous Database or Darktrace’s autonomous cyber-defense platform are based on AI and are, to a degree, capable of automated mitigation of various security problems, but they are still dependent on their respective teams of experts ready to intervene if something does not go as planned. That’s why such solutions are only offered as a part of a managed service package – even the best “autonomous AIs” still need humans from time to time…
So, is Artificial Intelligence the solution for all current and future cybersecurity challenges? Perhaps, but please do not let over-expectations or fears affect your purchase decisions. Thanks to the ongoing developments both in narrow and general AI, we already have much better security tools than just several years before. Yet, when planning your future security strategy, you still must think in terms of risks and the capabilities needed to mitigate them, not in terms of technologies.
Also, don’t forget that cybercriminals can use AI to create better malware, too. In fact, things are just starting to get interesting!