TL;DR
This paper introduces a novel approach to detect malicious browser extensions by decomposing their runtime behavior into "aspects" and clustering these aspects. By building profiles of these aspect clusters for both extensions and their categories, the system can effectively identify suspicious extensions that deviate from the norm, offering a more granular and robust security analysis.
What Problem Does This Paper Address?
Browser extensions offer powerful functionality but also pose significant security risks. Millions of users download extensions from web stores without adequate scrutiny, making it difficult to distinguish benign from malicious ones. Existing security measures often struggle with issues like code obfuscation, the need for browser modification, and high overhead. Specifically, extensions can exploit their privileged access to breach sandboxing policies, steal user data, or even masquerade as legitimate tools. The paper aims to bridge the gap between developers, web stores, and users by providing a reliable way to understand and assess extension behavior before installation.
Methods
The core of the proposed solution is aspect-level behavior decomposition and clustering.
- Behavior Representation: The system uses System Call Dependence Graphs (SCDGs) to represent an extension's runtime behavior. SCDGs capture the interactions between a program and the operating system by representing system calls as vertices and their dependencies as edges.
- Aspect Decomposition: To achieve a more granular analysis, the SCDGs are further decomposed into Aspects of Extension Behavior (AEBs). An AEB is essentially a meaningful subgraph of an SCDG that represents a specific functional aspect of the extension's behavior (e.g., "bookmark accessing," "DOM storage accessing").
- Clustering: Similar AEBs from different extensions are grouped into "AEB clusters" using subgraph isomorphism (specifically, y-isomorphism for efficiency).
- Profiling: Each extension is then represented by a vector of AEB clusters, forming its "extension profile." Similarly, categories of extensions (as organized in web stores) are represented by "category profiles," which are the union of the profiles of extensions within that category.
- Detection: Suspicious extensions are identified by comparing their extension profile against the profile of their assigned category. If an extension's profile is not a subset of its category's profile, it's flagged as suspicious, indicating the presence of outlier AEBs.
- Dynamic Tracing: To capture runtime behavior, the system employs dynamic tracing, addressing challenges like the input space issue (using dynamic symbolic execution) and differentiating system calls between the browser and extensions (using Callgrind and trace differentiation).
Datasets
The paper evaluates its approach using a dataset of over 1,000 Firefox extensions.
- Training Set: 1,107 extensions were used for training, encompassing 8 categories.
- Testing Set: 186 extensions were used for testing, including 8 known malicious extensions and 1 malicious extension created by the authors. The paper does not introduce a new benchmark dataset but rather utilizes existing extensions from web stores.
Key Insights and Ideas
- Granularity Matters: Decomposing behavior into "aspects" (AEBs) provides a more nuanced understanding than looking at the entire system call trace, allowing for finer-grained detection of anomalies.
- Category Profiles as Baselines: Leveraging the fact that extensions within the same category tend to share similar functionalities and thus exhibit similar behavioral aspects, category profiles serve as effective baselines for identifying deviations.
- Uniqueness and Exclusiveness: The detection logic relies on the idea that legitimate extensions within a category will have profiles that are subsets of the category's profile. Suspicious extensions, by exhibiting unique or exclusive behaviors, will break this subset relationship.
- Dynamic Symbolic Execution for Input Space Coverage: Addressing the challenge of exploring all possible execution paths of an extension is crucial. The use of dynamic symbolic execution helps to significantly increase the coverage of execution paths and inputs.
- System Call Differentiation is Key: Accurately distinguishing system calls made by the extension from those made by the host browser is vital for accurate behavior analysis.
Key Results
- Effectiveness in Detection: The system successfully detected all 9 malicious extensions in the testing set, achieving a 0% false negative rate.
- Input Space Improvement: The use of the input resolver (IR) led to an average 54.8% increase in the total number of SCDGs, indicating better exploration of execution paths. Outliers in SCDGs after using the IR were minimal (average 0.4%).
- Suspicious AEB Clusters: On average, 6.0% of suspicious AEB clusters were identified in the testing set, indicating deviations from category profiles.
- Efficiency: The system demonstrates reasonable performance, with the three main components (Dynamic Tracing, SCDG Extracting, SCDG Clustering) consuming an average of 116.7 seconds per category. Dynamic tracing accounted for the largest portion (48.4%).
- Scalability: The approach scales well, handling over 1,000 extensions, which is a significant improvement over previous studies that typically analyzed around 20 extensions.
- False Positive Analysis: A false positive rate of approximately 10% was observed in the testing set, primarily due to extensions belonging to multiple categories. The authors propose combining category profiles for multi-category extensions to mitigate this.
Limitations and Future Work
- System Call Tracing Limitations: The approach is not applicable if an extension invokes no system calls, though this is rare for malicious extensions.
- Mimicry Attacks: While difficult, attackers could potentially mimic behaviors by carefully crafting system calls and object codes. The use of y-isomorphism in clustering helps mitigate this by considering dependencies.
- JavaScript Injection: The current system does not explicitly track JavaScript injected into web pages by extensions. Future work could involve modifying the system to identify and track such injections.
- Multi-Category Extensions: The current detection mechanism might flag extensions belonging to multiple categories as suspicious. The authors suggest combining category profiles for such extensions to improve accuracy.
Summary updated