Internet traffic is now routinely stored as "hypersparse" matrices. These are matrices indexed by every possible address (for IPv4, $2^{32}$ rows by $2^{32}$ columns), of which only a vanishing fraction of entries are nonzero. A software standard called GraphBLAS turns graph algorithms into sparse linear algebra on these matrices, and pipelines built on it ingest $10^8$ to $10^9$ packets per second. Studies of tens of trillions of packets built this way have found that traffic obeys very stable heavy-tailed laws. What they measure, though, is one-dimensional: how many links or packets each address has. Separately, hyperbolic embedding places network nodes in negatively curved space, where room grows exponentially with radius. It is known to represent heavy-tailed, tree-like networks faithfully in as few as two dimensions.
The obvious combination is to "embed traffic hyperbolically and detect anomalies." It hides a trap. In hyperbolic network models, a node's distance from the origin is essentially the logarithm of its degree, so the radial half of the embedding restates what the heavy-tail statistics already capture. The only possible new information lies in the angle: which addresses sit near each other on the circle. This note proposes a GraphBLAS-native two-dimensional hyperbolic embedder for windowed traffic matrices. We use it to answer a falsifiable question: does the angular coordinate of real traffic carry stable, predictive structure that neither degree statistics nor the IP prefix tree already explain? The answer could be no, and the experiments are designed so that "no" is a clean, publishable finding.
Core Claim
We claim that angular position in a 2D hyperbolic embedding of a traffic window captures who-talks-to-whom structure: coordinated scanning campaigns, shared service clusters, and target sets. This structure is invisible to degree-only models. It persists across windows in benign backbone traffic, and it distinguishes coordinated darkspace sources. We further claim that this embedding can be computed with sparse matrix-vector products over the same partitioned matrices the existing pipelines already build. It would then add a modest constant factor to ingestion cost rather than requiring a separate machine-learning system.
The claim differs from the closest existing work in three ways. Anomaly and Change Detection in Graph Streams through Constant-Curvature Manifold Embeddings embeds each small graph in a stream as a single point and tests for drift. We instead embed the nodes of one enormous, evolving matrix. HC-GLAD trains hyperbolic contrastive networks on small graph-classification benchmarks, and it cannot run anywhere near line rate. NetSMF shows that sparse factorization scales embedding to 67.7 million nodes, but it does so in flat space and on static graphs. None of the three isolates the angle-versus-degree question, which we treat as the actual scientific content.
Research Program
Embedding primitive. The method follows the lesson of Bridging ML and algorithms: comparison of hyperbolic embeddings. That work found combinatorial embedders such as BFKL roughly 100 times faster than Riemannian gradient descent, at equal or better quality. It also found, through its description-length metric (Information Control Value, ICV), that 2D embeddings are often more robust than high-dimensional ones.
We assign each node's radius in closed form from its degree in the window. Angles are then refined by an angular smoothing step. We represent each node's angle as a unit complex number $z_v = e^{i\theta_v}$ and update
$$z \leftarrow \mathrm{normalize}\big(A_w\, (z \odot g(r))\big),$$
where $A_w$ is the symmetrized traffic matrix for window $w$, $g(r)$ down-weights high-degree hubs according to their radius, $\odot$ is elementwise multiplication, and normalize projects each entry back to the unit circle. Each iteration is a single complex-valued sparse matrix-vector product, so it is expressible in GraphBLAS.
Because the aggregation is a plain sum, it inherits the sum-partitioning property $A = A_1 \oplus \dots \oplus A_P$ from GraphBLAS Mathematical Opportunities. Partial products computed on separate cores or sensors combine by addition before normalization. Streaming windows, built as $A_t + A_{t+dt}$, warm-start from the previous window's angles.
Permutation behavior. Anonymization with a prefix-preserving hash is a row and column permutation, as noted in Anonymized Network Sensing Graph Challenge. Our embedding therefore commutes with anonymization. Embeddings computed on anonymized matrices can be shared between organizations in the style of DBOS Network Sensing without exposing raw addresses.
Scientific analysis. The analysis compares the angular structure against two explicit null explanations. The first is degree alone. The second is the IP prefix hierarchy, which the prefix-preserving hash retains and which is itself a tree that hyperbolic space would embed trivially. We also test the tension raised in the file (Gap 4): whether a 2D disk is enough for a $2^{32}$-address matrix, or whether ICV favors more dimensions. We answer it by comparing 2D, 3D, and higher-dimensional variants under ICV rather than under the ranking metric (Mean Average Precision, MAP), which the bake-off shows is biased toward high dimension.
Baselines
Degree-only statistical model. The naive reference point is the two-parameter Zipf-Mandelbrot degree model of What is Normal? and Large Scale Enrichment and Statistical Cyber Characterization of Network Traffic. We extend it to link prediction through a degree-product (configuration-model) score, and to anomaly detection through deviation of fitted parameters. It runs on the same windowed GraphBLAS matrices. Beating it means higher next-window MAP and detection of campaigns it provably cannot see.
Radial-only ablation. This is our own embedding with angles randomized, keeping only degree-derived radius. It isolates the key ingredient. If the full model does not beat it, the angle carries nothing.
Prefix-tree baseline. Hyperbolic distance is computed from the anonymized /8, /16, /24 prefix tree. It is the strongest "free hierarchy" competitor, because it requires no learning and survives anonymization.
Euclidean sparse factorization. This is NetSMF-style randomized SVD at dimension 128 applied to the same windows, tuned over the same window sizes. On static graphs, NetSMF reports embedding OAG (67.7M nodes, 895M edges) in 24 hours and beating DeepWalk and node2vec by up to 34% Micro-F1 (NetSMF). There is no reported result on Lacuna for NetSMF on traffic matrices.
Strongest existing hyperbolic embedder. This is BFKL itself, run offline on the same snapshots. On YEAST it reports MAP 0.756 against 0.542 for the Lorentz gradient-descent model (Bridging ML and algorithms). There is no reported result on Lacuna for any hyperbolic embedder on traffic data.
Our streaming GraphBLAS embedder should match BFKL's MAP within noise while running inside the ingestion pipeline. Beating BFKL on quality is not required; matching it at streaming cost is the claim. All methods receive identical windows, the same held-out windows, and equal hyperparameter-search budgets.
Why not just…?
Why not just fit Zipf-Mandelbrot degree statistics, which already run at $10^9$ packets per second? This alternative is the real threat. The radial coordinate is roughly log-degree, so on any task driven by activity volume the degree model should tie us. The experiments settle this with degree-preserving injected campaigns. These are coordinated sources whose degree sequence is statistically identical to background, so the degree model is blind to them by construction. The radial-only ablation then measures whether angle adds anything on real data. If angle adds nothing on real traffic, we report that traffic geometry is exhausted by its degree law.
Why not just use Euclidean randomized SVD on the sparse matrix? It is mature and GraphBLAS-compatible. The bake-off found that 200D Euclidean embeddings did worse than low-dimensional hyperbolic ones. Zambon et al. report change-detection rates of 0.04 for Euclidean against 0.96 for hyperbolic embeddings at Difficulty 2 on Delaunay streams. Neither result is on traffic, so we run the comparison directly, at matched memory: 2D hyperbolic embeddings need two numbers per node, and 128D Euclidean embeddings need 128.
Why not just use the IP prefix tree as the hierarchy? Prefix-preserving anonymization keeps it, and it costs nothing. If learned angles mostly reproduce prefix neighborhoods, the embedding is redundant. We measure the correlation between angular distance and prefix distance. We also test whether the embedding retains predictive power after conditioning on prefix distance.
Why not just count shared targets exactly with GraphBLAS? Co-targeting, computed as $A A^\top$, is exact but produces a dense, quadratically growing matrix among active sources. The embedding compresses it to one angle per node. We compare detection quality against exact co-targeting on windows small enough to compute it. This bounds how much the compression loses.
Proposed Experiments
Link prediction on MAWI. Benign sources in MAWI persist for years. On MAWI windows ranging from $2^{17}$ to $2^{27}$ packets (the range over which Focusing and Calibration confirmed heavy-tail stability), we predict which source-destination links recur in the next window. We report MAP and ICV for all baselines.
Angular stability. We align embeddings from windows weeks apart by the best rotation of the circle, a one-parameter search. We then measure how well angular neighborhoods are preserved for sources that persist. This mirrors the stability results reported for degree laws, and it asks whether "normal" traffic also has a stable geometry.
Coordinated darkspace sources. On CAIDA telescope data joined to GreyNoise enrichment, we test whether angularly close sources share tags, such as EternalBlue scanning over TCP/445, more often than degree-matched or prefix-matched pairs. Malicious sources in this data have half-lives of hours to days, which stresses the warm-started streaming update.
Injected campaigns. We insert degree-preserving coordinated campaigns into MAWI windows and measure detection rate as the campaign fraction shrinks. This is the one setting where the degree baseline is guaranteed to fail and exact co-targeting provides an upper bound.
Cost. We report throughput of the full pipeline with and without the embedding step on the synthetic PCG32 benchmark and on the 246 GB enterprise capture. The reference pipeline processes $10^8$ to $10^9$ packets per second, and the enterprise capture was previously processed 3x faster than real time.
What decides the question. The idea succeeds if full embeddings beat the radial-only and prefix baselines on held-out MAWI MAP and on CAIDA tag co-occurrence, detect injected campaigns the degree model misses, match BFKL quality, and slow ingestion by less than a small constant factor. It fails informatively if the angle collapses onto degree or prefix structure.
References
- Anomaly and Change Detection in Graph Streams through Constant-Curvature Manifold Embeddings
- HC-GLAD: Dual hyperbolic contrastive learning for unsupervised graph-level anomaly detection
- NetSMF: Large-Scale Network Embedding as Sparse Matrix Factorization
- Bridging ML and algorithms: comparison of hyperbolic embeddings
- GraphBLAS Mathematical Opportunities: Parallel Hypersparse, Matrix Based Graph Streaming, and Complex-Index Matrices
- Anonymized Network Sensing Graph Challenge
- DBOS Network Sensing: A Web Services Approach to Collaborative Awareness
- What is Normal? A Big Data Observational Science Model of Anonymized Internet Traffic
- Large Scale Enrichment and Statistical Cyber Characterization of Network Traffic
- Focusing and Calibration of Large Scale Network Sensors Using GraphBLAS Anonymized Hypersparse Matrices