TL;DR
This paper analyzes how three major malware delivery operations (Dridex, Dorkbot, and Dyre-Upatre) evolved over a year, specifically in response to law enforcement takedown attempts. It reveals that these operations exhibit both predictable (displacement, defiance) and unpredictable behaviors, highlighting the need for adaptive and multi-faceted disruption strategies.
What Problem Does This Paper Address?
The cybersecurity landscape is constantly threatened by sophisticated malware and botnet operations. Law enforcement agencies (LEAs) and security companies frequently conduct "takedown" operations to disrupt these malicious activities. However, the success of these takedowns is often mixed, and there's a significant knowledge gap regarding how these operations adapt and evolve after being targeted. This paper aims to fill that gap by providing a detailed, longitudinal study of how specific malware delivery operations change their infrastructure, tactics, and behaviors in the wake of disruption attempts. Understanding these evolutionary patterns is crucial for developing more effective and sustainable countermeasures.
Methods
The researchers developed a novel methodology to track and analyze malware delivery operations over time. This involved:
- Building Download Graphs: Constructing directed graphs where nodes represent entities like file hashes (SHA-2), URLs, IP addresses, and fully qualified domain names (FQDNs). Edges represent the relationships between these entities as observed in download events.
- File Classification: Labeling each file node as malware, potentially unwanted program (PUP), or benign, using ground truth sources like VirusTotal, AVClass, and the National Software Reference Library (NSRL). They also aggregated family aliases to ensure consistent labeling.
- Tracking Operational Activity: Developing metrics to quantify different aspects of malware delivery operations, including:
- Network Dynamics: Metrics related to the infrastructure used (URLs, FQDNs, IPs, countries), and how they are associated (e.g., IPs per FQDN, indicating Fast Flux).
- Downloader Dynamics: Metrics related to the binaries themselves, such as download counts, dropping events, file churn (how often new variants appear), file sizes, and detection scores (malice and prevalence).
- Time Series Analysis: Analyzing these metrics over a year to observe trends and changes in behavior, particularly around the times of takedown operations.
Datasets
The primary dataset used in this study is download activity metadata provided by Symantec. This anonymized dataset comprises download data from 12 million end-users who opted into a data-sharing program. It includes information on:
- The time of download.
- The server from which the binary was downloaded.
- The program that initiated the download.
- Metadata about the downloaded binaries, such as filename, size, and Symantec's "reputation" and "prevalence" scores.
This dataset covers the period from October 1st, 2015, to September 29th, 2016.
Key Insights and Ideas
- Predictable vs. Unpredictable Responses: Malware operations don't just disappear after a takedown. They exhibit both predictable reactions like displacement (moving infrastructure) and defiance (increasing activity) and less predictable, novel behaviors.
- The "Super Binary" Phenomenon: A small number of malware binaries are responsible for a disproportionately large amount of download activity. Identifying and targeting these "super binaries" could be a highly effective strategy.
- Distributed Architectures are Key: All studied operations relied heavily on distributed delivery infrastructures (CDNs, shared hosting), making them resilient to single points of failure and difficult to take down completely.
- Polymorphism is Rampant: Malware constantly churns out new variants (high SHA-2 churn), making detection a continuous challenge. Dorkbot, in particular, showed extremely high churn.
- Takedowns Can Cause Shifts: Takedowns can force malware operators to change their infrastructure, sometimes shifting from distributed email services to more centralized DGA (Domain Generation Algorithm) servers, or vice-versa.
- PUPs and Malware Intertwined: The Upatre operation, for instance, was observed dropping a significant amount of Potentially Unwanted Programs (PUPs) alongside malware, highlighting the complex ecosystem of malicious software.
Key Results
- Dridex: Showed increased and diversified network activity during its DNS sinkhole takedown, followed by a decrease. It heavily utilized distributed architectures and showed bursts of dropping activity, which was uncharacteristic for this financial fraud trojan.
- Dorkbot: Exhibited highly cyclic and stochastic activity, with a sharp but brief drop post-takedown, but no long-term observable effects. It relied on upstream droppers and showed a high proportion of files engaging in dropping behavior, with extremely high polymorphism and low detection rates.
- Dyre-Upatre: After the Dyre takedown, Upatre showed a significant shift in infrastructure, moving from email services to DGA servers and a CDN. It dropped a range of software, primarily PUPs. Its activity showed a peak before the takedown and then a decrease, but with a notable increase in upstream droppers in the latter half of the year.
Limitations and Future Work
- Data Limitations: The study is limited by the focus on only one stage of the malware supply chain (software download) and potential limitations in VirusTotal's coverage.
- Ground Truth Scarcity: Detailed ground truth on takedown operations is often scarce, making it difficult to precisely correlate observed behaviors with specific intervention details.
- Survivorship Bias: The studied botnets were targeted, meaning they might have characteristics that made them more likely to be disrupted than others, potentially limiting the generalizability of findings to untargeted operations.
- Future Work: The authors suggest using causal inference frameworks to better analyze takedown effects, incorporating more diverse data sources (e.g., network traffic from ISPs, dark web discussions) for a more holistic view, and exploring ecosystem dynamics to understand how operators replace each other's infrastructure. They also emphasize the need for better coordination between LEAs, security companies, and service providers.
Summary updated