DRL Deep Reinforcement Learning for Better Cybersecurity Defences

Needs usually diminish, and that’s the way it goes the majority of the time for whatever reason. But as so much more of the work and personal worlds for people has gone digital and ever greater amount of everything is in the Cloud there is so much opportunity out there for cyber attackers to go after and attempt to acquire valuable data and information. From malware to ransomware and all wares in between, they’re out there and they’re becoming more complex right in step with how the digital world makes its own daily advances.

Here at 4GoodHosting like any other good Canadian web hosting provider we have hosting SSL certificates that can secure a website for basic e-commerce purposes. But that’s the extent of what folks like us are able to offer with regards to web security. Cybersecurity is a much lager umbrella, and a more daunting one if it’s possible for an umbrella to be daunting. But fortunately there are much bigger players at work working on defences so the good guys still have a chance of staying intact in the face of ever-great cybersecurity threats.

One of the more promising developments there as of recently is Deep Reinforcement Learning, which is an offshoot of sorts from other artificial intelligence aims where researchers found cross-purpose applications for what they’d been working with. So let’s use this week’s blog entry to look at this as these days nearly every one has some degree of an interest in cybersecurity. If not an outright need for it.

Smarter & More Preemptive

Deep reinforcement learning offers smarter cybersecurity, the ability for earlier detection of changes in the cyber landscape, and the opportunity to take preemptive steps to scuttle a cyber attack. Recent and thorough testing with realistic and widespread threats had deep reinforcement learning being effective at stopping cyber threats and rendering them inept up to 95% of the time. The performance of deep reinforcement learning algorithms is definitely promising.

It is emerging as a powerful decision-support tool for cybersecurity experts and one that has the ability to learn, adapt to quickly changing circumstances, and make decisions autonomously. In comparison to other forms of AI that will detect intrusions or filter spam messages, deep reinforcement learning expands defenders’ abilities to orchestrate sequential decision-making plans so that defensive moves against cyberattacks are undertaken more ‘on the fly’ and in more immediate response to threats that are changing as they happen.

This technology has been built with the understanding that an effective AI agent for cybersecurity needs to sense, perceive, act and adapt, based on the information it can gather and on the results of decisions that it enacts. Deep reinforcement learning has been crafted with that need taken very much into account, combining reinforcement learning and deep learning to that it is entirely agile and adept in situations where a series of decisions in a complex environment need to be made.

Incorporating Positive Reinforcement

Another noteworthy functionality of DRL is how good decisions leading to desirable results are reinforced with a positive reward that is encompassed as a numeric value, and then at the same time bad choices leading to undesirable outcomes come with a negative cost. This part of DRL has strong fundamental A.I. underpinnings as it is similar to how people learn tasks. Children at a young age learn that if they do something well that leads to a favorable outcome as seen that way by people expecting it of them, they know they will benefit from that in some way.

The same thing of sorts occurs with DLR here in deciphering cybersecurity threats and then disabling them. The agent can choose from a set of actions. With each action comes feedback, good or bad, that becomes part of its memory. There’s an interplay between exploring new opportunities and exploiting past experiences and working through it all builds memory as to what works well and what doesn’t.

4 Primary Algorithms

Recent advances with DLR that have taken it to the next level and put it on the radar for the cybersecurity world as a promising new A.I.-based technology have been based on four deep reinforcement learning algorithms  – DQN (Deep Q-Network) and three variations of what’s known as the actor-critic approach. Here is an overview of what was seen in the trials:

  • Least sophisticated attacks: DQN stopped 79% of attacks midway through attack stages and 93% by the final stage
  • Moderately sophisticated attacks: DQN stopped 82% of attacks midway and 95% by the final stage
  • Most sophisticated attacks: DQN stopped 57% of attacks midway and 84% by the final stage. This was notable as it was far higher than the other 3 algorithms

While DRL for cybersecurity looks promising and may someday be a well-known acronym in the world of web technology and online business, the reality is that for now at least it will need to be working in conjunction with humans. A.I. can be good at defending against a specific strategy but isn’t as adept with understanding all the approaches an adversary might take and it is not ready to completely usurp human A.I. cybersecurity analysts yet.

Post Navigation