TL;DR
This paper proposes a novel approach to predict future network intrusions by leveraging Hidden Markov Models (HMMs) and alert clustering. By incorporating contextual information like source/destination IPs and alert types, their method aims to provide more actionable intelligence for intrusion response systems compared to existing methods that focus solely on intrusion type or category.
What Problem Does This Paper Address?
Network security is a critical concern, and while intrusion detection systems (IDS) are good at identifying ongoing or past attacks, they are reactive. The next logical step is to predict future intrusions to enable proactive defense. Current intrusion prediction methods often fall short by focusing too narrowly on just the type or category of an attack. This leaves responders with incomplete information, making it difficult to craft effective countermeasures. Furthermore, many existing methods rely on specific domain knowledge or scenario-specific information, limiting their adaptability. The authors aim to bridge this gap by developing a prediction method that is adaptable, doesn't require deep domain expertise, and provides richer contextual information for better response planning.
Methods
The proposed framework consists of three main modules:
- Alert Pre-Processing: Raw alerts from intrusion detection systems are standardized into a common format, and duplicate alerts are removed based on key attributes (timestamp, IP addresses, ports, signature, category).
- Alert Clustering (Bag-of-Words + k-means): To create a sequence of symbols for the HMM, alerts are first represented using a Bag-of-Words (BoW) model. Important attributes like source/destination IPs, ports, alert type, and alert category are extracted. IP addresses are treated specially, with each segment considered a separate "word." This vocabulary is then used with the k-means clustering algorithm to group similar alerts into distinct clusters. Each cluster represents a unique "symbol" for the HMM.
- Alert Prediction (Hidden Markov Model): A Hidden Markov Model (HMM) is trained on the sequence of cluster IDs generated from historical alerts. The trained HMM is then used to predict the next alert cluster in a sequence. The Viterbi algorithm is used to find the most likely sequence of hidden states, and then the model predicts the next observation (alert cluster) based on state transition and emission probabilities. The prediction can be for a single step or multiple steps ahead.
Datasets
The primary dataset used for training and evaluation is the DARPA dataset, specifically the LLDOS 1.0 and LLDOS 2.0.2 multi-stage attack datasets. The paper notes that DARPA 2000 LLDOS 1.0 was used for vocabulary generation and clustering, while DARPA 2000 LLDOS 2.0.2 was used as the test network for monitoring and prediction.
Key Insights and Ideas
- Context is King: The paper strongly emphasizes that predicting just the "what" (type/category) of an intrusion isn't enough. Including "who" (source IP) and "where" (destination IP range) provides crucial context for effective response.
- Adaptability through Training: By relying on a training process rather than hardcoded domain knowledge, the HMM approach becomes adaptable to different network environments and evolving attack patterns.
- Clustering as a Symbol Generator: The novel use of BoW and k-means to transform complex alert data into a sequence of discrete cluster IDs is key to making it amenable to HMM modeling. This effectively creates a "language" for network attacks.
- Trade-offs in Clustering: The number of clusters (symbols) for the HMM presents a trade-off. More clusters can lead to better separation of unique alert types but might increase the complexity and potentially reduce accuracy if the number of states in the HMM isn't sufficient. Fewer clusters might merge distinct alert types, hindering prediction.
Key Results
- Alert Cluster Prediction: Achieved accuracies of 81% for single-step prediction and 77% for five-step prediction of the next alert cluster.
- Alert Category Prediction: Achieved accuracies of 95% for single-step prediction and 92% for five-step prediction of the next alert category.
- Comparison to Prior Work: The proposed method showed a 5% improvement in alert category prediction accuracy over existing variable-length Markov chain methods.
- Impact of Hidden States: The optimal number of hidden states for the HMM was found to be around 8 for alert cluster prediction, yielding a Level 1 accuracy of 67%.
- Impact of Training Length: Longer training sequences (up to 3500 alerts) significantly improved prediction accuracy, reaching 73% for Level 1.
- Impact of Cluster Size: A smaller number of clusters (e.g., 5) generally led to higher prediction accuracy (up to 88% for Level 1), though it could hinder the separation of distinct alert types.
Limitations and Future Work
- Unseen Intrusion Types: A significant limitation is that the current model cannot predict intrusion types that were not present in the training data.
- False Alerts and Deception: Identifying false alerts and sophisticated attacker deception tactics remains a challenge.
- Scalability of Clustering: While the HMM itself is adaptable, the BoW and k-means clustering might face scalability issues with extremely large and diverse datasets.
- Future Work: The authors suggest increasing prediction accuracy with larger cluster sizes, improving the ability to predict unseen intrusion types, and developing methods to identify false alerts and misleading attacker actions. They also note that the number of hidden states in the HMM should be carefully chosen, as too few states can hinder the modeling of complex state changes.
Summary updated