Abstract
Advanced Persistent Threat (APT) attack usually refers to the form of long-term, covert and sustained attack on specific targets, with an adversary using advanced attack techniques to destroy the key facilities of an organization. APT attacks have caused serious security threats and massive financial loss worldwide. Academics and industry thereby have proposed a series of solutions to detect APT attacks, such as dynamic/static code analysis, traffic detection, sandbox technology, endpoint detection and response (EDR), etc. However, existing defenses are failed to accurately and effectively defend against the current APT attacks that exhibit strong persistent, stealthy, diverse and dynamic characteristics due to the weak data source integrity, large data processing overhead and poor real-time performance in the process of real-world scenarios. To overcome these difficulties, in this paper we propose APTSHIELD, a stable, efficient and real-time APT detection system for Linux hosts.
Introduction
Advanced Persistent Threats (APTs) are the "ninjas" of the cyber world: patient, well-funded, and capable of remaining hidden for months. Traditional security tools often fail because they either rely on known signatures—which APTs easily bypass—or generate a massive volume of log data that is impossible to process in real-time.
Existing Endpoint Detection and Response (EDR) solutions face three major hurdles:
- Data Integrity: Attackers often tamper with user-level logs to obscure their activities.
- The "Dependency Explosion": Tracking system interactions over long periods creates massive provenance graphs that can crash a system's memory.
- Real-time Lag: Analyzing complex attack chains often happens "post-mortem" rather than during the active phase of an attack.
System Architecture
To address these challenges, APTSHIELD utilizes a three-layered architecture designed for stability, efficiency, and real-time responsiveness on Linux hosts.
The system consists of three primary modules:
- Data Collection Module: Stably collects kernel-level data to ensure data source integrity.
- Data Compaction Module: Reduces data volume through redundant semantics skipping and non-viable node pruning.
- APT Detection Framework: Maps system activities to the MITRE ATT&CK framework using a label-based tracking mechanism.
Stable Data Collection and Compaction
A critical component of APTSHIELD is the selection of a robust data source. After benchmarking several tools including Sysdig, Lttng, and Auditbeat, Auditd was selected for its superior stability and low overhead.
As shown in the performance comparison, Auditd maintains remarkably low CPU and memory usage even under full-load conditions, making it the "gold standard" for enterprise-grade monitoring. To prevent memory bloat, APTSHIELD employs two novel pruning techniques: Redundant Semantics Skipping, which ignores repeated identical actions, and Non-viable Node Pruning, which removes "dead-end" processes that exit without performing suspicious actions.
Label-based Detection Framework
Instead of resource-intensive graph matching, APTSHIELD utilizes a Label Transfer mechanism. Entities (processes and files) are assigned "status" and "behavior" labels based on the MITRE ATT&CK framework. As processes interact, these labels flow and aggregate. When a specific combination of labels (e.g., "Network Access" + "Sensitive File Read" + "Shell Execution") triggers a rule, the system fires an alert in real-time. This allows the system to maintain the context of an 80-day attack without keeping 80 days of raw logs in RAM.
Key Results
- Efficiency: APTSHIELD saved between 1,000ms and 11,000ms of processing time per 1 million events compared to non-compacted systems.
- Low Overhead: Pruning reduced the memory footprint by up to 22.6 MB, maintaining average CPU usage between 4-8% under load.
- Detection Performance: The system successfully detected all tested APT scenarios (Webshells, RATs, and file-less attacks) with a low false-positive rate, outperforming state-of-the-art tools like Sleuth and Holmes.
Conclusion and Future Work
APTSHIELD demonstrates that real-time APT detection is achievable by prioritizing context over raw content. By tracking the "flow of untrustedness" via labels and aggressively pruning redundant data, the system maintains high-fidelity detection over long-term attacks. Future work will focus on mitigating "taint explosion" and expanding rule sets using machine learning to automatically discover new suspicious label flow patterns.
Summary updated