The speed of AI progress is accelerating at breakneck speed. This year, we saw some very cool industry breakthroughs with AI - and we’re excited to share them with you.
AI’s war of algorithms
The objective of Artificial Intelligence is to enhance the ability of machines to process copious amounts of data and by doing so, automate a broad range of tasks. Despite this benign objective, AI also lends itself to nefarious ends, and in our increasingly digitising world, AI has the potential to cause an unprecedented degree of damage. In the same way that human intelligence can be used towards positive, benign or detrimental purposes, so can artificial intelligence.
Throughout 2019, our research team has perceived a potential war of algorithms, where good AI will be forced to contend with bad AI. Recently released research has shown that AI has the potential to be used in three different ways; in the business logic of the attack, within the infrastructure framework of an attack or in an adversarial approach, to undermine AI based security systems. With the theoretical groundwork already established, the cyber-attack landscape is at the precipice of becoming vastly more sophisticated and complex. Armed with this powerful technology hackers can become more robust, and we will soon be facing attacks that are more devastating in their capability and impact. The need for a cybersecurity paradigm shift has never been greater.
Most solutions available today are woefully under-prepared to deal with these huge operational challenges. They can’t adequately fight against complex AI attacks because they employ sophisticated evasion techniques that hide algorithms capable of more severe damage.
In 2020, organisations need to enter this new era fully aware of this impending threat and ensure the ongoing security of their data and systems with a solution that is up to the task.
The transition beyond EDR
Fortunately, AI technologies are advancing, and deep learning (the most advanced form of AI) is proving to be the most effective cybersecurity solution for threat prevention. Deep learning is inspired by the brain’s ability to learn new information and from that knowledge, predict accurate responses. Once a brain learns to identify an object, its ongoing identification becomes second nature. Similarly, it has been discovered that as the artificial deep neural network brain learns to identify any type of cyber threat, its prediction capabilities become instinctive. For enterprises, this has significant implications as it means any kind of malware, known and unknown, are predicted and prevented with unmatched accuracy and speed.
Unlike detection and response-based solutions (which wait for the attack to execute before reacting) the deep learning neural network enables the analysis of files pre-execution so that malicious files can be prevented pre-emptively. By taking a preventative approach, files and vectors are automatically analysed statically prior to execution. This is critical in a threat landscape, where real time can sometimes be too late.
The expansion of deep learning models
During 2019, one of the major trends in AI was how the size of deep learning models kept growing at an accelerating pace. This was very exciting because it meant that larger sets of data that are comprised of greater complexity can now be processed. For example, state-of-the-art language translation models used at the end of 2019 were many times larger than those used at the end of 2018. The result being that instead of paying attention to sentence combinations as the basis of data sets, the model is now learning in more granular detail and assigning meaning to smaller word combinations. This trend of growing the layers of deep learning models is expected to develop at an exponential pace. This trend is also underscoring the importance of growing computational efforts and the cost required in training state-of-the-art models.
The ability to manipulate machine learning classifiers
In recent years, adversarial learning, the ability to fool machine learning classifiers using algorithmic techniques has become a hot research topic. However, this past year has seen a diffusion of such research from the limited domain of image recognition to other, more critical domains, particularly the ability to bypass cybersecurity next generation anti-virus products. In July, a cyber-research company Skylight discovered that they were successfully able to undermine the machine learning algorithm of a leading cybersecurity product. By carefully analysing the engine and model of the product, they were able to identify a particular bias towards a specific pattern, from which they were then able to craft a simple bypass by appending a selected list of strings to a malicious file.
Exploration of generative adversarial networks
2019 saw several mergers and acquisitions of smaller companies and more strategic big investments in technologies that can cross platforms and protect against different and future attack vectors. It is unlikely that this is going to slow down or stop. There is still room for innovation - in fact, one area that is particularly interesting is Generative Adversarial Networks (GAN). In May 2019, researchers at Samsung demonstrated a GAN-based system that produced videos of a person speaking with only a single photo of that person provided. Then in August of this year, a large dataset consisting of 12,197 MIDI songs each with their own lyrics and melodies were created through neural melody generation from lyrics by using conditional GAN-LSTM.
As 2019 proved to be a landmark year in both cybersecurity and artificial intelligence, 2020 shows no signs of things slowing down as new threats continue to arise daily. With this in mind, enterprises of all sizes should continue to keep their eyes peeled while ensuring their respective organisations are fully protected with the latest threat prevention solutions to keep themselves and their data fully protected – with AI and deep learning at the front lines.
Guy Caspi, CEO, Deep Instinct