TL;DR
This paper introduces a novel deep learning-based system for detecting Advanced Persistent Threats (APTs) by analyzing DNS activity. It addresses the limitations of previous methods by incorporating features that capture the relationship between DNS requests and responses, achieving high accuracy in identifying suspicious behaviors.
What Problem Does This Paper Address?
Advanced Persistent Threats (APTs) are a significant concern, especially in sensor networks, due to their stealthy, long-duration, and evolving nature. Traditional detection methods struggle to keep up with these sophisticated attacks. While analyzing DNS requests has shown promise, existing approaches often overlook the crucial interplay between DNS requests and their corresponding responses. This oversight can lead to missed detections, particularly in APT attack scenarios where the DNS resolution process might be incomplete or manipulated. The paper aims to bridge this gap by developing a more comprehensive detection system that leverages the full DNS transaction.
Methods
The core of this research is a deep learning-based detection framework. The system comprises four main units:
- DNS Data Collector: Gathers DNS request and response logs from the network.
- DNS Data Pre-Processing: Filters and prepares the raw DNS data. This involves a preliminary feature generator and a pre-processing filter that removes irrelevant data and applies basic checks.
- DNS Behavior Feature Extraction: This is a key innovation. The authors define eight features to characterize DNS behavior, with three being entirely new. Crucially, they introduce features that capture the relationship between DNS request and response messages, as well as temporal characteristics of DNS logs.
- DNS Behavior Analysis and Evaluation: A deep neural network (specifically, a 6-layer fully connected network) is employed to analyze the extracted features and assign a "suspiciousness" score to DNS behaviors.
The deep learning model uses the rectified linear unit (ReLU) activation function and a cross-entropy loss function for training.
Datasets
The study utilizes a large dataset of DNS request records collected from the Jilin University Education Network over 47 days.
- Initial Data: 4,907,147,146 DNS request records.
- After Pre-processing: 376,605,606 records.
- Black Data (Simulated APT Traffic): 15,338 records, generated by simulating various APT attack techniques described in reports.
- White Data (Benign Traffic): Derived from the top one million domain names based on Alexa rankings, assumed to be innocent.
- Grey Data: The remaining pre-processed data after filtering for white-listed domains and visit frequency.
The paper notes that obtaining real APT attack data is difficult, leading to the use of simulated "black data."
Key Insights and Ideas
- The Request-Response Relationship is Crucial: The paper strongly emphasizes that analyzing only request or response messages in isolation is insufficient. The relationship between them provides vital clues for detecting sophisticated attacks like APTs, especially those employing DNS tunneling.
- Temporal Features Matter: The timing and frequency of DNS resolutions, as well as the intervals between requests and responses, are significant indicators of malicious activity. APTs often exhibit low-frequency, discontinuous traffic patterns.
- Deep Learning for Complex Patterns: The authors argue that deep learning is superior to traditional machine learning for this task due to its ability to model complex relationships between features and handle large datasets effectively.
- Feature Engineering is Key: The identification and creation of new features, particularly those capturing the request-response dynamics and temporal aspects, are central to the system's effectiveness.
Key Results
The proposed deep learning-based system demonstrates strong performance in detecting suspicious DNS behavior associated with APT attacks:
- Accuracy: Achieves an average accuracy of 97.6%.
- Recall: Reports a recall of 96.8%, indicating a high ability to identify actual malicious activities.
- False Positive Rate (FPR): Maintains a low FPR of 2.3%, minimizing the misclassification of benign traffic as malicious.
The paper also shows that the newly introduced features significantly improve detection performance compared to using older feature sets.
Limitations and Future Work
The authors acknowledge several limitations and suggest future directions:
- Over-sampling Challenges: Due to the difficulty in obtaining real APT data, the "black data" is simulated, and over-sampling techniques were used, which can introduce noise.
- False Positives from Shared Infrastructure: The system might flag some normal DNS traffic as suspicious if it shares a DNS server with malicious domains or exhibits similar resolution times. This is identified as an optimization goal.
- Feedback Mechanism: The authors propose using a feedback mechanism to upgrade the detection system by identifying suspicious domain names and hosts and then using this information to refine the detection process.
- Real-World APT Data: While the system was validated with simulated data, applying it to more real-world APT scenarios is a desired future step, though obtaining such data remains a challenge.
Summary updated