Criminal Neural Networks: Can Hackers Turn Artificial Intelligence to Evil?

Our world is changing, and new technologies bring not only new opportunities but also new dangers. This article focuses on neural networks—not on what they can do by themselves (which is widely covered elsewhere), but on the potential harm neural networks could cause in the hands of hackers.

The Image of the Hacker and the Role of AI

In modern media, hackers are often portrayed as wizards or miracle workers, wielding scripts, viruses, and other digital “spells.” These criminal hackers are seen as evil magicians, using cutting-edge tools for malicious purposes. One such tool, of course, is artificial intelligence.

What Is Artificial Intelligence?

Artificial intelligence (AI) is an umbrella term that includes a variety of tools related to neural networks, big data, and machine learning. Most often, it refers to a neural network trained to solve a specific, usually interpretive, task: speech recognition, image analysis, contextual text translation, and so on.

A typical neural network is, in essence, a complex mathematical function f(x, a) that depends on a set of inputs x and parameters a. It receives data in a specific format and produces a result.

Building a simple neural network involves defining the task, choosing the network type, and finding a labeled dataset. The network is then trained—its parameters a are automatically adjusted so that the function produces the desired results on the training data. After testing, if all goes well, the network is ready for use—assuming the dataset was good, the architecture was appropriate, and nothing broke along the way.

It’s important to note that neural networks are highly specialized tools designed for specific tasks. Attempts to retrain a network or expand its application often result in it not only failing to learn the new task but also “forgetting” how to perform the original one. This narrow specialization appears to be a mathematical property, not dependent on the physical implementation. So, if you need to solve a new task—even a similar one—you’ll likely have to develop a new network from scratch.

Also, neural networks aren’t programmed in the traditional sense. This means verifying the theoretical correctness of their results (like checking an algorithm in a program) requires new methods and tools. In other words, ensuring the safety and reliability of neural networks is an unsolved problem that still needs appropriate solutions.

What Do Modern Hackers Do?

Today’s hackers have little in common with the lone professionals depicted in TV shows. Most cybercriminals are part of organized groups, often made up of many programmers of average skill. Their main goal, like any criminal organization, is to make a profit with minimal cost and risk. This shapes the main areas of cybercrime: the vast majority of their crimes involve stealing personal data, fraud, and extortion.

There are five main types of cyber threats:

1. Malware infections: Over half of all cyber threats involve infecting computers with malicious software. Beyond traditional viruses, hackers now embed malicious code in open-source projects or infect servers that distribute automatic software updates. Such attacks have infected tens or even hundreds of thousands of computers worldwide.
2. Individual scam attacks: These involve direct contact with users to trick them into giving up data or installing malware. Often, real people pose as support staff or bank employees. These attacks usually aim for direct theft of money, or they may result in ransomware infections that demand payment to unlock data.
3. Classic system hacking: Using programs and scripts to break into systems accounts for just under 17% of all cyber threats. This is often part of a larger scheme, such as infecting servers with malware. Cryptocurrency exchanges and platforms are frequent targets, along with traditional government institutions.
4. Website attacks: Online stores and other sites that collect personal information are popular targets. Data is either stolen or used for extortion.
5. DDos attacks: These never go out of style. July 2019 marked the 20th anniversary of the first DDos attack. The main idea is to create a botnet—computers ready to flood a server with requests on command, overloading it.

Do Hackers Need Neural Networks?

The media often highlights the potential dangers of neural networks in the hands of hackers. It’s believed that neural networks could make existing threats more dangerous in several ways:

• AI-powered social engineering: Neural networks could imitate human communication (e.g., via email) well enough to convince victims to transfer money or reveal private data.
• “Human-like” DDos attacks: While many systems can now detect and block traditional DDos attacks, neural network bots could mimic real user behavior, making attacks harder to spot.
• Bypassing antivirus protection: Neural networks might help malware evade detection.

All three scenarios are highly unlikely and rather far-fetched. For example, if neural networks capable of realistic conversation are developed, marketers will likely use them for automated marketing before hackers do. Similarly, neural networks that mimic user behavior are more valuable for simulating high website traffic than for crashing sites. As for neural networks that can bypass security, that’s a distant and uncertain future.

There are, however, more exotic scenarios. In 2018, hospitals and clinics were targeted by hackers. In one case, hospital management had to pay a ransom in bitcoin to regain access to patient data. Medical institutions often lag behind in IT security, using software that experts consider vulnerable to various attacks.

Fantasies vs. Reality

Some Israeli researchers explored the maximum damage hackers armed with neural networks could inflict on a medical facility. They imagined a scenario where hackers, exploiting vulnerable software, gain access to patient data, including X-rays. Suppose hackers use a neural network trained for a very specific task: removing cancer tumors from X-ray images. Such tampering, the researchers say, could have fatal consequences for patients.

Many articles like this appear, but they often overlook several factors. First, the neural network described would require a large team of specialists and access to data from specialized medical research. It’s unlikely that an average hacker group would have access to such training data. Second, the scenario involves a very expensive and specialized attack. Interestingly, the article only mentions altering X-rays, not other records like lab results, suggesting the authors are oversimplifying the complexity of such an attack.

It’s no surprise, then, that most experts agree: current neural networks are not suitable for typical hacker tasks due to the high complexity and cost of training. If new threats involving neural networks do arise, it will be because of new ways to use them.

So, Do Hackers Need Neural Networks?

Malware can already cause significant damage to industries. For example, WannaCry—a virus not originally intended for industrial systems—halted production at several Renault factories and affected Nissan’s manufacturing systems.

In short, the answer is twofold:

• Neural networks are not needed by ordinary hackers, but they may find use (or already be in use) in industrial espionage and state-level attacks.
• Technology is always advancing. While working with neural networks is currently time-consuming and expensive, that may not always be the case. The emergence of standardized libraries and open-access labeled datasets could change everything. Crimes involving neural networks could become part of everyday life, so it’s important to start thinking about these threats now.