Apposcopy: automated detection of Android malware (invited talk)

Yu Feng, Isil Dillig, Saswat Anand, Alex AikenDeMobile@SIGSOFT FSE 2014 · DOI

TL;DR

This paper introduces Apposcopy, a novel semantics-based approach for detecting Android malware that steals private information. By combining high-level, semantic malware signatures with a powerful static analysis, Apposcopy effectively identifies malicious applications with high accuracy, even against obfuscated variants.

What Problem Does This Paper Address?

The Android platform, with its vast user base, has become a prime target for mobile malware. A significant portion of this malware focuses on exfiltrating sensitive user data, such as personal information stored on the device. Existing detection methods, like taint analysis, struggle to distinguish malicious data leaks from legitimate app functionality, requiring significant manual effort. Signature-based detectors, on the other hand, often rely on low-level code patterns that are easily bypassed through simple obfuscation techniques, necessitating constant signature updates. This creates a need for a more robust and automated approach to detect Android malware that targets private information.

Methods

Apposcopy tackles this problem by introducing a two-pronged approach:

  • High-Level Malware Signatures: Apposcopy utilizes a Datalog-based specification language that allows security analysts to describe the semantic characteristics of malware families. This goes beyond simple code patterns and focuses on the intended behavior. These signatures can describe two types of properties:
  • Control-flow properties: These capture how different components of an Android application interact, for example, a broadcast receiver launching a service in response to a system event.
  • Data-flow properties: These describe the movement of sensitive data, such as reading device identifiers and sending them over the internet.
  • Powerful Static Analysis: To match these semantic signatures, Apposcopy employs a sophisticated static analysis framework that includes:
  • Inter-Component Call Graph (ICCG): This is a novel, high-level representation of Android applications that models inter-component communication. It's used to verify control-flow properties specified in the signatures.
  • Static Taint Analysis: This analysis tracks the flow of sensitive data from designated sources (e.g., device IDs) to designated sinks (e.g., network connections), enabling the verification of data-flow properties.

The combination of semantic signatures and deep static analysis makes Apposcopy resistant to low-level code transformations and obfuscation techniques.

Datasets

The paper evaluates Apposcopy on the following datasets:

  • Android Malware Genome project: A corpus of known Android malware samples.
  • Google Play apps: A dataset of legitimate applications from the official Android app store.
  • Obfuscated malware: Existing malware samples that have been intentionally obfuscated using tools like ProGuard.

Key Insights and Ideas

  • Semantics over Syntax: The core insight is that focusing on the meaning and behavior of malware, rather than just its code structure, leads to more robust detection. Semantic signatures are harder for malware authors to evade.
  • Bridging the Gap: Apposcopy effectively combines the strengths of taint analysis (identifying sensitive data flows) and signature-based detection (providing a structured way to define malicious behavior) into a unified framework.
  • High-Level Specifications are Practical: The Datalog-based language allows analysts to express complex malware behaviors in a concise and understandable manner, making it a practical tool for security professionals.
  • ICCG as a Key Abstraction: The inter-component call graph provides a crucial abstraction for analyzing the complex interactions between different parts of an Android application, which is essential for understanding control flow.

Key Results

  • High Accuracy on Known Malware: Apposcopy achieved an overall accuracy of 90.0% on a corpus of 1027 malware samples from the Android Malware Genome project, with a false negative rate of approximately 10.0% and a false positive ratio of less than 0.2%.
  • Effective Differentiation of Benign Apps: When evaluated on 11,215 Google Play apps, Apposcopy flagged only 16 as potential malware, all of which were confirmed as malicious via VirusTotal.
  • Resilience to Obfuscation: Apposcopy demonstrated a 100.0% detection rate on obfuscated malware, significantly outperforming other commercial anti-virus tools whose detection rates ranged from 14.3% to 57.1%.

Limitations and Future Work

  • Limitations: The paper acknowledges that while Apposcopy is effective, the accuracy of detection is dependent on the quality and completeness of the malware signatures provided by the analyst. The static analysis, while powerful, might still have limitations in handling extremely complex or dynamically generated code.
  • Future Work: The authors suggest that future work could involve developing more automated methods for generating semantic malware signatures, exploring dynamic analysis techniques to complement static analysis, and expanding the framework to detect other types of Android threats beyond information-stealing malware.

Summary updated

BibTeX
@inproceedings{feng2014apposcopy,
  author = {Yu Feng and Isil Dillig and Saswat Anand and Alex Aiken},
  title = {{Apposcopy: automated detection of Android malware (invited talk)}},
  booktitle = {DeMobile@SIGSOFT FSE},
  year = {2014},
  doi = {10.1145/2661694.2661697},
}
Figure