Abstract
This paper introduces a model-driven Decision Support System (DSS) that utilizes Bayesian Belief Networks (BBN) to counter passive reconnaissance in cybersecurity. By modeling how attackers "fingerprint" operating systems (OS) through passive network observation, the system identifies the minimum number of network features to modify to misdirect an attacker's perception. Experimental results show that the BBN model achieves predictive accuracy comparable to artificial neural networks while providing the transparency necessary for defenders to make informed strategic decisions.
Introduction
In cybersecurity, passive reconnaissance is a critical phase where attackers observe network traffic without direct interaction to identify a host's operating system and version. This process, known as OS fingerprinting, allows attackers to tailor exploits to specific vulnerabilities. Traditional defenses like firewalls often fail to prevent this silent observation. While cyber deception—making a system appear as a different OS—is a known strategy, many existing methods lack a sophisticated model of attacker belief formation. This research addresses this gap by using a probabilistic model to actively manipulate an attacker's perception.
Methodology
The proposed Decision Support System (DSS) consists of three main modules: a Database Module for traffic storage, a Model Management Module for probabilistic learning, and a Dialog Module for "what-if" analysis. The core of the system is a Bayesian Belief Network (BBN) that learns the relationships between network features, such as TCP window size, TTL, and ICMP packet characteristics, and the host OS.
The system employs a deception algorithm that utilizes the BBN’s Conditional Probability Tables (CPT) to select a minimal set of features to modify. By targeting the most informative features (e.g., ttl or winsize), the defender can achieve high levels of obfuscation with minimal network overhead, ensuring the system remains functional while appearing as a different, "fake" OS.
Experimental Results
The authors evaluated the BBN model against a baseline Sequential Artificial Neural Network (ANN) using passively collected traffic from various operating systems, including Linux (Ubuntu), Mac/iOS, Windows (7, 8, Vista), and Android. The BBN performed comparably to the ANN in terms of prediction accuracy across dataset sizes ranging from 9,000 to 168,000 datapoints. This demonstrates that interpretability—a key advantage of BBNs over "black box" ANNs—does not require a sacrifice in performance.
In a case study designed to make a Linux system appear as a Mac, the deception algorithm significantly reduced the identification metrics for the true OS. As shown in the evaluation matrices, the precision, recall, and F1-score for correctly identifying the host dropped sharply after the deception was applied, effectively blinding automated fingerprinting tools.
Conclusion
This research highlights that effective cyber deception is not about changing every packet header, but about changing the right ones. By leveraging the interpretability of Bayesian models, defenders can identify and modify the features that provide the highest information gain to an attacker. Future work will focus on generalizing this framework to other types of reconnaissance and ensuring real-time performance in production environments.
Summary updated