Abstract
Cloud storage provides external data storage services by combining and coordinating different types of devices in a network to work collectively. However, there is always a trust relationship between users and service providers, therefore, an effective security auditing of cloud data and operational processes is necessary. We propose a trusted cloud framework based on a Cloud Accountability Life Cycle (CALC). We suggest that auditing provenance data in cloud servers is a practical and efficient method to log data, being relatively stable and easy to collect type of provenance data. Furthermore, we suggest a scheme based on user behaviour (UB) by analysing the log data from cloud servers. We present a description of rules for a UB operating system log, and put forward an association rule mining algorithm based on the Long Sequence Frequent Pattern (LSFP) to extract the UB. Finally, the results of our experiment prove that our solution can be implemented to track and forensically inspect the data leakage in an efficient manner for cloud security auditing.
Introduction
Cloud storage relies on a fundamental "trust gap": users hand over their data to service providers but have limited visibility into how that data is actually handled. While security models like the Cloud Accountability Life Cycle (CALC) exist, they often struggle with the sheer volume of data traffic. Traditional auditing methods usually look at input/output data or simple logs, which are difficult to interpret at scale.
The challenge is twofold: first, how to transform low-level system calls (such as read, write, and open) into meaningful descriptions of user behavior; and second, how to perform this analysis efficiently enough to handle "Big Data" without the computational complexity inherent in standard data mining techniques.
The Cloud Accountability Life Cycle (CALC)
To address these challenges, this research utilizes the Cloud Accountability Life Cycle (CALC) as a foundational framework for trusted cloud services. This cycle ensures that security is not a one-time check but a continuous process of planning, sensing, logging, and auditing.
The cycle begins with Policy Planning and moves through Sense & Trace, Logging, and Safe-keeping of Logs. The latter stages involve Reporting & Replaying, Auditing, and finally Optimizing & Rectifying to close the loop.
System Architecture
The proposed Internal Provenance Cloud Security Auditing System is structured into three primary layers, influenced by external factors such as legal provisions and organizational policies.
- System Data Mining Layer: Collects raw data from the cloud network and file systems.
- Data Storage and Protection Layer: Focuses on maintaining data provenance and consistency, ensuring the integrity of the collected logs.
- Application Layer: Provides the interface for auditing and reporting based on the mined insights.
The LSFP Algorithm: Mining User Behavior
The core innovation of this work is the Long Sequence Frequent Pattern (LSFP) algorithm. Unlike standard algorithms like FP-growth that use tree structures to find frequent items, LSFP employs a Directed Graph.
Key Innovations:
- Graph-Based Representation: In this graph, "edges" represent the transfer relationship from one action to the next, preserving the chronological order of operations. This is crucial because in cloud auditing, the order of events often defines the behavior.
- Two-Stage Process:
- Training Phase: Scans a subset of data to build an incidence matrix and prune infrequent patterns based on a "minimum support" ($min\_sup$) threshold.
- Mining Phase: Divides long sequences of system activity into short, recognizable "User Behaviors" (e.g., "File Creation" or "File Edit").
- Provenance Integration: Provenance data—a history of every action taken on a file—is more stable and easier to collect than standard logs, making it a reliable source for auditing.
Experimental Evaluation
The researchers evaluated the LSFP approach using Linux operating system logs in a virtualized environment (CentOS 6.4 on VMware) to simulate a cloud node. Training sets ranged from 1M to 10M records.
Key Results:
- Linear Scalability: The algorithm's running time shows a substantially linear relationship with the size of the training set. Increasing data from 1M to 10M saw running times rise predictably from 400ms to 900ms.
- Rule Discovery: The number of discovered rules follows a non-linear logarithmic relationship with the training set size, suggesting diminishing returns in discovering new behavior types after a certain volume.
- Sensitivity Control: The $min\_sup$ parameter acts as a sensitivity knob. Lowering the threshold increases the number of identified behaviors but may introduce noise.
Conclusion and Future Work
By converting cryptic OS logs into readable "User Behaviors," this framework provides "Cloud Credibility"—auditing reports that are meaningful to human stakeholders and legal authorities. While the LSFP algorithm is efficient, future work will focus on automating the selection of the "minimum support" threshold and testing the framework across distributed, multi-tenant cloud environments to further refine forensic depth in data leakage scenarios.
Summary updated