Deep reinforcement learning, a form of autonomous artificial intelligence, proved effective in preventing 95 percent of cyber attacks in a study conducted by the Pacific Northwest National Laboratory.
The results show promise in the use of smarter cybersecurity in proactive cyberdefense, PNNL said.
MITRE’s ATT&CK framework was reportedly employed to develop the DRL algorithms, while the Open AI Gym helped in creating the attack simulation environment.
An algorithm called Deep Q-Network outperformed three variations of actor-critic approach algorithms in training cyberdefense agents on simulated attacks.
DQN stopped 79 percent of least sophisticated attacks midway through the process, and 95 percent during the final stage of the intrusion. In sophisticated attacks, DQN blocked the process midway through in 57 percent of the cases and in 85 percent during the final stage.
The findings were presented at the annual meeting of the Association for the Advancement of Artificial Intelligence in Washington, D.C. this month.