Autonomous Hardware-Based Proactive Defenses with Deep Reinforcement Learning
摘要
Traditional cybersecurity, with its reactive measures and static defenses, struggles to keep pace with increasingly sophisticated and persistent cyberattacks. Attackers hold a significant advantage, exploiting even a single vulnerability to compromise a system. They readily adapt to conventional defenses, leaving system administrators constantly struggling to react effectively. To overcome this asymmetry, a paradigm shift towards proactive and adaptive security is essential. This article introduces a novel hardware-based defense framework driven by a deep reinforcement learning (DRL) agent. This system anticipates, mitigates, and responds to cyber threats in real-time. Our design seamlessly integrates defender hardware structures into a CPU’s back-end pipeline stages. These structures intercept critical instructions, like system calls, during unknown application execution and neutralize malicious behavior using lightweight, user-defined defense subroutines. Furthermore, our DRL agent continuously monitors the behavior of the running application. By extracting relevant features from the application, the agent proactively anticipates threats and dynamically invokes hardware-based defenses. Experimental results demonstrate the efficacy of our hardware support against prominent ransomware samples, highlighting the efficiency of defense operations with a minimal performance impact of