Introduction
In cyber forensics, analysts frequently face the "needle-in-a-haystack" challenge: identifying a specific group of actors within a massive, anonymized dataset. The VAST Challenge 2020 Mini-Challenge 1 exemplified this problem, presenting a fictional global community graph with over 100 million edges and 200,000 vertices. The objective was to find a small "template" subgraph representing a known group's behavior within this noisy environment. To address this, we present CA2 (Cyber Attacks Analytics), a visual analytics system that combines coordinated multiple views with a human-in-the-loop iterative matching workflow.
System Architecture
CA2 is a web-based system built with React and D3.js, designed to leverage human intuition in guiding graph-matching algorithms. The interface is organized into three primary panels:
- Control Panel: Allows users to manage data input and filter communication "channels," such as emails, phone calls, or financial transactions.
- Organization Panel: Provides high-level views of group behavior, including temporal scatterplots (Time View), geographic distributions (Space View), and node-link diagrams (Structure View).
- Personnel Panel: Offers detailed individual profiles and a heatmap for comparing activity frequencies across potential candidates.
Iterative Matching Workflow
The core of CA2 is an Iterative Matching Workflow. Rather than attempting to match the entire 100-million-edge graph at once—a task prone to noise and computational overhead—the system facilitates a "seed-and-grow" approach.
Analysts first identify a "seed" (a small, unique set of matched nodes) based on rare behavioral signatures. The system then employs a five-step algorithm to retrieve and suggest adjacent nodes. The human analyst remains in the loop, verifying these suggestions across multiple dimensions before officially adding them to the matched set.
Behavioral Verification via Coordinated Views
A critical insight of the CA2 system is that structural matches (who interacts with whom) must be validated by behavioral matches (when and where they interact). By linking the Structure View with Space and Time views, analysts can quickly discard false positives that appear structurally identical but exhibit different activity patterns.
For instance, while communication data (emails/phones) is often dense and noisy, "procurement edges" (buying/selling) proved to be rare and highly distinctive. In the VAST 2020 dataset, a unique signature involving two individuals linked by a procurement edge to a single item served as the anchor for the entire search.
Results
The CA2 system successfully identified the target group within the VAST Challenge 2020 dataset. The iterative workflow effectively extracted a subgraph from the massive database that closely mirrored the provided template. This streamlined approach to complex graph analysis earned the authors an Honorable Mention for Streamlined Analysis Process at the competition, demonstrating the efficacy of combining automated suggestions with expert human verification.
Summary updated