Program Behavior Dynamic Trust Measurement and Evaluation Based on Data Analysis

Shuai Wang, Aiqun Hu, Tao Li, Shaofan LinSymmetry 2024 · DOI

TL;DR

This paper introduces a novel dynamic trust measurement system for industrial control terminals, crucial for detecting sophisticated attacks like code reuse and mimicry. By analyzing program behavior through function call sequences and system calls, their method achieves higher detection accuracy and speed compared to existing techniques.

What Problem Does This Paper Address?

Industrial control systems are critical infrastructure, but their terminals are vulnerable to attacks that can disrupt production and cause significant economic losses. Traditional security methods often lack the granularity and real-time responsiveness needed to detect modern threats like code reuse and mimicry attacks. These attacks can manipulate program execution flow or disguise malicious actions as legitimate operations, making them difficult to identify with coarse-grained security measures. Existing detection techniques for these specific attacks also suffer from drawbacks like high latency, complexity, or limited accuracy, necessitating a more robust and efficient solution.

Methods

The paper proposes a two-pronged approach to dynamic trust measurement:

  • Function Call Sequence Analysis:
    • Benchmark Library Construction: They build a "benchmark library" by analyzing the program's control and data dependencies to extract function call sequences. This involves constructing an Abstract Syntax Tree (AST), Control Flow Graph (CFG), Program Dependency Graph (PDG), and finally a Function-Level Execution Flow Graph (FEFG).
  • Sliding Window Measurement: A sliding window technique is employed to measure function call sequences in real-time. This allows for partial tracking and comparison against the benchmark library, enabling a more timely evaluation than full sequence analysis.
  • Dynamic Credibility Evaluation: Real-time function call subsequences are matched against the benchmark library. A mismatch triggers an "untrustworthy" flag, potentially terminating the program. Successful matches are recorded to aid subsequent comparisons.
  • System Call Sequence Analysis:
    • Partitioning: Program behavior is divided into function call behavior and system call behavior within function intervals.
  • Maximum Entropy Model: For system calls, a maximum entropy model is used. This statistical approach is chosen because system call sequences can be variable and uncertain. The model leverages known information to infer the most uniform probability distribution, effectively measuring the credibility of system call sequences within function intervals.
  • Feature Extraction and Training: Features like call frequency and time are extracted, and the maximum entropy model is trained on this data to predict system call behavior.

Datasets

The paper mentions using the "VX Heaven [30]" experimental sample database for evaluation. This database appears to be a collection of malicious software samples and their associated behaviors, used to test the effectiveness of the proposed model against various attack types.

Key Insights and Ideas

  • Partitioning is Key: Dividing program behavior into function call sequences and system calls within function intervals is crucial for managing complexity and improving detection.
  • Symmetry in Evaluation: The concept of "symmetry" is central, aiming for a balanced comparison between the established benchmark library and real-time measurements.
  • Maximum Entropy for Uncertainty: Applying maximum entropy to system call analysis is a smart move, as system calls are inherently less predictable than function calls, and this method effectively handles probabilistic uncertainty.
  • Sliding Window for Real-time: The sliding window approach for function call sequences is a practical solution for achieving real-time monitoring without the overhead of analyzing entire execution paths.

Key Results

  • Improved Detection Accuracy: The proposed method demonstrates higher accuracy in detecting ROP attacks (40% higher than C-FLAT), mimicry attacks (48% higher), and code injection attacks (10% higher).
  • Faster Detection: The system achieves higher detection efficiency due to the sliding window and benchmark library approach, combined with the probabilistic evaluation of system calls.
  • Lower Latency: Experimental results show that the proposed system has a lower overall latency compared to an n-gram based scheme, making it more suitable for real-time industrial control scenarios.
  • Better Unbiasedness: The model exhibits better unbiasedness, as indicated by lower MAPE values compared to other tested models like Var-gram, Dyck, and FSA.

Limitations and Future Work

  • Operating System Specificity: The current implementation is for the Linux operating system. Future work should focus on extending its portability to multiple operating systems.
  • Model Training Time: The maximum entropy model requires multiple iterations with historical data, which can be time-consuming and data-intensive. Further optimization for improved self-learning efficiency is needed.
  • Mimicry Attack Limitations: While better than C-FLAT, the paper notes that C-FLAT can partially detect mimicry attacks, implying that even the proposed method might have nuances in completely preventing all forms of this attack.

Summary updated

BibTeX
@article{wang2024program,
  author = {Shuai Wang and Aiqun Hu and Tao Li and Shaofan Lin},
  title = {{Program Behavior Dynamic Trust Measurement and Evaluation Based on Data Analysis}},
  journal = {Symmetry},
  year = {2024},
  doi = {10.3390/sym16020249},
}
Figure