Abstract: A new multi-stage attack correlation method based on attack sequence is proposed in this paper. The algorithm first excavates the attack sequence of network attack behaviors from a large number of security events, and then analyzes the correlation of the events which are in accord with certain attack pattern using membership function. The simulation results show that the algorithm can not only correlate multiple isolated security events in attack scenarios to detect composite attack, but also can find the real security threat hidden in security events.
Introduction
Modern enterprise networks are protected by a diverse array of security devices, including firewalls, Intrusion Detection Systems (IDS), and vulnerability scanners. While these tools are essential, they often operate in isolation, generating a massive volume of uncoordinated alerts. This leads to "information overload," where security analysts struggle to identify genuine threats amidst a sea of false positives. Furthermore, sophisticated cyberattacks are rarely single events; they typically occur in multiple stages—such as scanning, exploitation, and the installation of backdoors. Isolated alerts fail to provide the "big picture" necessary to recognize these coordinated campaigns.
To address this, the paper introduces a multi-stage correlation algorithm designed to mitigate alarm fatigue by identifying recurring attack sequences in network traffic. By combining chronological sequence mining with fuzzy causal analysis, the method successfully reconstructs complex attacks from thousands of fragmented security alerts.
Methodology
The proposed framework transforms raw security events into actionable intelligence through a four-step pipeline:
- Global Sequence Construction: All security events are sorted chronologically and labeled by event type to create a master timeline of the network's security state.
- Candidate Generation: The global stream is divided into smaller "candidate sequences" based on a sliding time window ($W_T$). This groups events that occur within a specific temporal proximity, suggesting a potential relationship.
- Maximal Sequence Mining: The algorithm identifies "maximal attack sequences"—frequently occurring patterns that are not merely subsets of other, longer sequences. This step utilizes a "minimum support" threshold to filter out coincidental or random event groupings.
- Fuzzy Causal Analysis: To confirm if events in a sequence are truly related, a membership function is applied. This function calculates a similarity score based on security attributes (such as IP addresses or ports) and assigns weights to these attributes. If the score exceeds a specific threshold ($T_\sigma$), a causal relationship is confirmed, linking isolated alerts into a single composite attack scenario.
Experimental Evaluation
The researchers evaluated the algorithm using the DARPA LLDOS 1.0 dataset from MIT Lincoln Lab. This benchmark contains a complete, five-stage attack scenario: IPSweep scanning, a sadmind exploit attempt, a sadmind buffer overflow, the installation of a DDOS Trojan via telnet, and finally, a stream DDOS attack against a victim host.
The simulation environment, as illustrated in the network topology above, demonstrates how various hosts and switches are monitored by IDS units to generate the raw event data required for correlation.
Key Results
In the DARPA LLDOS test, the algorithm demonstrated significant data reduction and filtering capabilities:
- Input: 4,216 original raw events.
- Processing: Reduced to 288 "senior" events and 262 candidate sequences.
- Output: Identified 94 attack sequences that met the support threshold.
- Success: The algorithm successfully extracted the core 5-step attack sequence with a support rate of 36%, effectively isolating the real threat from the background noise of the network.
Discussion and Insights
Two primary insights emerge from this research. First, cyberattacks follow a logical "grammar" (reconnaissance $\rightarrow$ exploitation $\rightarrow$ objective). By treating security events as a sequence rather than isolated points, the algorithm models the attacker's intent. Second, the use of fuzzy logic is crucial for real-world environments. Traditional rigid rules fail when data is missing or inconsistent; a membership function allows the system to handle "fuzzy" relationships where attributes might not match perfectly but still indicate a high probability of a connection.
Limitations and Future Work
While effective, the algorithm's performance depends heavily on the manual calibration of the time window ($W_T$) and support thresholds. If set too strictly, the system may miss "low and slow" attacks that unfold over extended periods. Future work will focus on automating the optimization of these parameters and benchmarking the algorithm's computational overhead in high-traffic, real-time environments.
Summary updated