Search

Modern cybersecurity best practices and standards require continuous Vulnerability Assessment (VA) and Penetration Test (PT). These activities are human- and time-expensive. The research community is trying to propose autonomous or semi-autonomous solutions based on Deep Reinforcement Learning (DRL) agents, but current proposals require further investigations. We observe that related literature reports performance tests of the proposed agents against a limited subset of the hosts used to train the models, thus raising questions on their applicability in realistic scenarios. The main contribution of this paper is to fill this gap by investigating the generalization capabilities of existing DRL agents to extend their VAPT operations to hosts that were not used in the training phase. To this purpose, we define a novel VAPT environment through which we devise multiple evaluation scenarios. While evidencing the limited capabilities of shallow RL approaches, we consider three state-of-the-art deep RL agents, namely Deep Q-Network (DQN), Proximal Policy Optimization (PPO), and Advantage Actor–Critic (A2C), and use them as bases for VAPT operations. The results show that the algorithm using A2C DRL agent outperforms the others because it is more adaptable to unknown hosts and converges faster. Our methodology can guide future researchers and practitioners in designing a new generation of semi-autonomous VAPT tools that are suitable for real-world contexts.

We present the first dataset that aims to serve as a benchmark to validate the resilience of botnet detectors against adversarial attacks. This dataset includes realistic adversarial samples that are generated by leveraging two widely used Deep Reinforcement Learning (DRL) techniques. These adversarial samples are proved to evade state of the art detectors based on Machine- and Deep-Learning algorithms. The initial corpus of malicious samples consists of network flows belonging to different botnet families presented in three public datasets containing real enterprise network traffic. We use these datasets to devise detectors capable of achieving state-of-the-art performance. We then train two DRL agents, based on Double Deep Q-Network and Deep Sarsa, to generate realistic adversarial samples: the goal is achieving misclassifications by performing small modifications to the initial malicious samples. These alterations involve the features that can be more realistically altered by an expert attacker, and do not compromise the underlying malicious logic of the original samples. Our dataset represents an important contribution to the cybersecurity research community as it is the first including thousands of automatically generated adversarial samples that are able to thwart state of the art classifiers with a high evasion rate. The adversarial samples are grouped by malware variant and provided in a CSV file format. Researchers can validate their defensive proposals by testing their detectors against the adversarial samples of the proposed dataset. Moreover, the analysis of these samples can pave the way to a deeper comprehension of adversarial attacks and to some sort of explainability of machine learning defensive algorithms. They can also support the definition of novel effective defensive techniques.

Adversarial attacks represent a critical issue that prevents the reliable integration of machine learning methods into cyber defense systems. Past work has shown that even proficient detectors are highly affected just by small perturbations to malicious samples, and that existing countermeasures are immature. We address this problem by presenting AppCon, an original approach to harden intrusion detectors against adversarial evasion attacks. Our proposal leverages the integration of ensemble learning to realistic network environments, by combining layers of detectors devoted to monitor the behavior of the applications employed by the organization. Our proposal is validated through extensive experiments performed in heterogeneous network settings simulating botnet detection scenarios, and consider detectors based on distinct machine- and deep-learning algorithms. The results demonstrate the effectiveness of AppCon in mitigating the dangerous threat of adversarial attacks in over 75% of the considered evasion attempts, while not being affected by the limitations of existing countermeasures, such as performance degradation in non-adversarial settings. For these reasons, our proposal represents a valuable contribution to the development of more secure cyber defense platforms.