Abstract
Traditional malware defenses are increasingly bypassed by code polymorphism and infrastructure agility. This paper introduces AMICO (Accurate Malware Identification via Classification of live network traffic Observations), a system designed to detect malicious software as it traverses the network. By shifting focus from content-based signatures to "provenance"—the behavioral footprint of a file's origin and download context—AMICO achieves high detection rates for zero-day threats. In a nine-month live deployment, the system demonstrated a 90% true positive rate with a 0.1% false positive rate, identifying significant gaps in traditional blacklisting and antivirus coverage.
Introduction
The security industry is currently locked in an arms race against malware distribution. Attackers employ two primary tactics to evade detection: code polymorphism, which frequently alters binary code to evade antivirus (AV) signatures, and infrastructure agility, which involves rotating domain names and IP addresses to bypass URL blacklists.
Traditional defenses are inherently reactive, requiring a threat to be observed elsewhere before a signature or blacklist entry can be generated. This creates a "detection gap" that allows new malware to infect thousands of users. AMICO addresses this by analyzing the context of a download in real-time, identifying malicious patterns in how files are delivered rather than relying on the files themselves.
The AMICO System Architecture
AMICO is designed to sit at the edge of a network, performing real-time reconstruction of HTTP traffic to identify Windows executable (PE) files. The system operates by extracting features related to the download's provenance and feeding them into a machine learning classifier.
As shown in the architecture above, the system reconstructs downloads from the network stream and maintains a download history. This history is used to learn provenance models that inform a Random Forest classifier. When a new download is detected, the Provenance Classifier evaluates it against these models to generate malware download alerts.
Methodology: Provenance-Based Detection
For every detected download, AMICO extracts a feature vector categorized into five groups:
- Past File Downloads: Tracks the frequency and stability of a file hash. Benign software (e.g., browser updates) is typically downloaded by many users, whereas polymorphic malware often appears unique to a few victims.
- Domain Features: Analyzes the reputation and history of the domain and its effective second-level domain (2LD).
- Server IP Features: Examines the server's IP and BGP prefix. While domains are easily changed, acquiring new IP ranges is more costly for attackers.
- URL Features: Inspects the structure of the URL (path, query strings) for anomalies common in malware distribution compared to legitimate mirrors.
- Download Request Features: Evaluates HTTP headers, such as the
Refererand the depth of the URL path.
Experimental Evaluation
The researchers deployed AMICO at the edge of a large academic network serving tens of thousands of users for nine months. The system was trained on two months of labeled data and tested against six months of live traffic, comprising nearly 43,000 total downloads. Ground truth was established using a "trusted" subset of nine major AV engines via VirusTotal.
Key Results
- High Accuracy: AMICO achieved a 90% True Positive Rate (TPR) with a remarkably low 0.1% False Positive Rate (FPR).
- Zero-Day Detection: The system successfully identified 78.6% of "zero-day" malware—samples that were initially flagged as benign by all nine trusted AV engines but later confirmed as malicious.
- Blacklist Limitations: A critical finding was that 97.5% of the malware caught by AMICO was not present in Google Safe Browsing (GSB) at the time of the download.
- Infrastructure Insights: The study found that 18% of malware downloads originated from the top 10,000 Alexa domains, proving that simple whitelisting of popular sites is an insufficient defense.
Conclusion and Future Work
AMICO demonstrates that while malware code is easy to change, building a reliable distribution infrastructure is a significant bottleneck for attackers. By exploiting the statistical "rarity" and suspicious footprints of these delivery networks, AMICO provides a robust defense against emerging threats.
Future work involves addressing the increasing prevalence of HTTPS, which would require AMICO to be integrated into a web proxy for visibility. Additionally, while AMICO focuses on the initial infection vector, expanding the system to detect post-infection command-and-control traffic remains a key area for development.
Summary updated