💻 computer science

Latent Geometry as a Structural Monitor: Eigenspace Alignment for Anomaly Detection in Anonymity Networks

This paper proposes a structural-monitoring framework that detects anomalies in the Tor anonymity network by analyzing eigenspace alignment within a stable nine-dimensional latent geometry, successfully identifying connectivity degradation events with zero false positives while falsifying traditional relay-departure hypotheses.

Original authors: Vaibhav Chhabra

Published 2026-05-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Vaibhav Chhabra

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the Tor network (a system designed to keep internet users anonymous) not as a list of computers, but as a giant, living city. This city has thousands of "relay" buildings that pass messages around. Usually, the city is stable. But sometimes, things go wrong: buildings might get damaged, traffic routes might get blocked, or a massive crowd might suddenly show up.

Traditional security systems are like burglar alarms. They wait until a window is broken or a door is forced open before they scream "Alert!" The problem is, by the time the alarm sounds, the damage is already done.

This paper proposes a different kind of system. Instead of waiting for a break-in, it acts like a structural engineer watching the city's foundation. It doesn't wait for a building to fall; it measures the pressure building up in the walls before anything breaks.

Here is how the paper explains this, using simple analogies:

1. The Two "Observers" (The Detective Team)

The system uses two different "detectives" to watch the city, and they speak different languages:

The Geometric Detective (CDAE): This detective looks at the shape of the city. Imagine the city is made of a flexible rubber sheet. This detective measures how much the sheet is stretching or bending. If the city is healthy, the rubber sheet bends easily (elastic). If something is wrong, the sheet gets stiff and starts to crack.
The Thermodynamic Detective (GRBM): This detective looks at the energy or "heat" of the city. Imagine the city has a mood. Is everyone calm, or is the population getting agitated and fragmented? This detective senses if the "mood" of the network is fracturing before the shape actually changes.

They are connected by a translator (CCA). This translator checks if both detectives agree on what they are seeing. If one sees a crack and the other sees calm, that disagreement itself tells a story about what kind of problem is happening.

2. The "Stiff" vs. "Soft" Directions

The paper introduces a cool concept: the city has "stiff" directions and "soft" directions.

Soft Directions: These are like the city's parks or empty streets. The population can move around here freely without causing stress.
Stiff Directions: These are like the city's load-bearing pillars. If the population is forced to move into these pillars, it means the structure is under extreme stress.

The system's main job is to watch for "Stiff-Axis Collisions." If the city's population suddenly tries to move into the "pillars" instead of the "parks," the system knows a major structural event is happening.

3. The "Mass" vs. "Gravity" Trick

Sometimes, the city gets a huge crowd of new people. Is this a good thing (a festival) or a bad thing (a flood of spies)?

Mass: This counts every single person equally.
Gravity: This counts only the "heavy" people (the ones carrying the most traffic).

The paper found a way to tell the difference. If the "Mass" moves a lot but "Gravity" stays put, it means a huge crowd of light, low-traffic people arrived (a population surge). If both move together, it means the core, heavy infrastructure is shifting (a structural fracture).

4. What They Actually Found

The researchers watched this "city" for 67 days. They didn't just guess; they tested their system against real events.

The Big Discovery (February 20, 2026): The system detected a massive structural stress signal. Usually, you'd think this meant a bunch of buildings (relays) collapsed. But the researchers checked the logs and found zero buildings had actually left.
- The Real Cause: It turned out to be a roadblock (a BGP routing issue with Cloudflare). The buildings were still there, but the roads connecting them were cut off. The system was smart enough to detect that the connectivity was broken, even though the buildings were still standing. This is a type of failure that normal alarms miss.
The "Fake" Surge (February 5–13): The system also saw a huge crowd arrive. But because the "Mass" moved while "Gravity" stayed still, the system correctly identified this as a flood of new, lightweight relays (likely a coordinated deployment by a single operator), rather than a structural collapse.

5. The Bottom Line

This paper claims to have built a structural monitor for the internet.

It doesn't need to know what an "attack" looks like in advance.
It doesn't need to see the actual messages being sent (which are hidden in Tor).
It simply measures the geometry and energy of the network's behavior.

If the network's "shape" starts to deform in a way that hits the "stiff pillars," the system raises an alarm. It successfully proved that you can detect a network failure (like a roadblock) before it causes a total collapse, simply by watching how the network's "body" moves.

In short: Instead of waiting for the house to burn down, this system feels the heat rising in the walls and tells you, "Something is wrong with the structure," even if the fire hasn't started yet.

Technical Summary: Latent Geometry as a Structural Monitor

Problem Statement

Traditional network anomaly detection relies on reactive, threshold-based mechanisms that identify events only after they manifest (e.g., when signals cross a predefined limit). This approach fails to detect the structural pressure that accumulates prior to a transition. In encrypted behavioral populations like the Tor anonymity network, where packet-level visibility is absent by design, standard monitoring often misses failure modes that do not involve explicit topology changes (e.g., relays leaving the consensus). The paper posits that while geometry (the measured state) changes, the underlying structure (the load-bearing organization of the population) precedes it. The challenge is to measure this structural deformation using only public relay metadata to detect network failures before they result in catastrophic threshold breaches.

Methodology

The authors propose a dual-observer pipeline that treats the relay population as a geometric energy landscape. The system operates on 67 consecutive daily observation windows (January 19 – April 21, 2026) using 191 behavioral features derived from the Onionoo API.

1. Dual-Observer Architecture

Geometric Observer (CDAE): A Contractive Denoising Autoencoder maps 17 selected continuous features (excluding binary flags to avoid Jacobian discontinuities) into a 32-dimensional latent space. The training loss includes a Frobenius norm penalty on the Jacobian ( $\|J\|_F^2$ ) to suppress sensitivity to noisy, high-variance directions, forcing the encoder to learn "stiff" (load-bearing) and "soft" (elastic) axes.
Thermodynamic Observer (GRBM): A Gaussian Restricted Boltzmann Machine operates on the full 191-feature space. To avoid sigmoid saturation caused by extreme bandwidth outliers, the system measures the visible free energy ( $F_{visible}$ ) directly from the visible layer before activation, rather than reading hidden layer activations. The Coefficient of Variation (CV) of this energy across the population serves as a fragmentation signal.
The Bridge (CCA): Canonical Correlation Analysis (CCA) bridges the CDAE and GRBM latent representations. It tracks the agreement strength ( $\rho_1$ ) and the rotation angle ( $\theta$ ) between the two observers' shared structural axes over time.

2. Eigenspace Alignment and Stiff Subspace

The core innovation is the Eigendecomposition of the Jacobian Trace (EJT). By decomposing the metric tensor $J^\top J$ , the system identifies:

Stiff Subspace ( $V_{stiff}$ ): Directions where the encoder is most sensitive and resists deformation.
Soft Subspace ( $V_{soft}$ ): Directions where the population can move freely (elastic absorption).
Invariant Structure: The system empirically determined that exactly $k=9$ dimensions are required to capture 90% of the trace mass of $J^\top J$ . This 9-dimensional subspace remained invariant across all 67 observation windows, validated at a 16.8 $\sigma$ separation from a Gaussian noise baseline.

3. Event Classification and Discriminators

The pipeline classifies events into six regimes based on gate activations:

PRECURSOR: Thermodynamic fragmentation (high CV) without geometric deformation.
REGIME_S (Elastic Stretch): Population surge absorbed elastically (high $\alpha_{guard}$ , high $\delta_{mg}$ ).
REGIME_D (Deformation): Localized geometric shift without thermodynamic fragmentation.
REGIME_E (Stiff Fracture): Global movement into stiff axes (low EJT z-score), indicating structural stress.
REGIME_K (Administrative Maintenance): Coordinated fleet restarts distinguished via forensic checklists.
NORMAL: No gates fire.

A critical discriminator, $\delta_{mg}$ (Mass-Gravity Divergence), distinguishes between a benign population surge (where unweighted "mass" shifts but traffic-weighted "gravity" remains stable) and genuine structural fractures.

Key Results

1. Structural Stability and Noise Floor

Null Baseline: Monte Carlo simulations with 1,000 iterations of Gaussian noise yielded a maximum CCA correlation of $\rho_1 = 0.614$ .
Empirical Signal: Real Tor data yielded $\rho_1 = 0.9992$ , confirming the structural agreement is a property of the network, not the pipeline architecture.
Stiff Subspace Invariance: The 9-dimensional stiff subspace was stable across all 67 windows. The top 10 features in this subspace showed a median Jaccard similarity of 0.90 between consecutive windows, dominated by geographic (latitude/longitude) and temporal (days since restart) features.

2. Confirmed Event: February 20, 2026

The pipeline detected a REGIME_E (Stiff Fracture) event on February 20, 2026, coinciding with a confirmed Cloudflare BYOIP configuration change that resulted in BGP prefix withdrawals.

Signals: Global EJT z-score = $-4.38$ (4.38 standard deviations below baseline); CCA rotation $\theta = 67.13^\circ$ .
Forensic Validation: The authors tested the "relay departure hypothesis" (that relays physically left the network). Cross-referencing 116 departed relays against Cloudflare prefixes revealed zero matches.
Conclusion: The event was a connectivity degradation without topology change. The network's effective routing paths were severed while relays remained listed in the consensus, a failure mode invisible to standard relay-count monitoring.

3. False Positive Rate (FPR)

Primary Gates: The Ch5 CV gate and Ch6 Global EJT gate achieved a 0.0% FPR on 24 confirmed stable windows.
Secondary Gates: The Ch6 Guard Elastic gate had an FPR of 14.3% (reducing to 3.6% when population surges were reclassified as investigated events). The Ch1 CCA rotation gate had a 49% FPR and is used only as a supporting signal.

4. Forensic Discoveries

February 05–13 Surge: Identified a coordinated deployment of 7,600 relays from residential IPs (USA/Argentina) by a single operator (printerexpert@mail2tor.com). The deployment was rejected by 72.9% of directory authorities. The $\delta_{mg}$ discriminator correctly identified this as an elastic population surge (REGIME_S) rather than a fracture.
April 07–08 Maintenance: A "stiff fracture" signal was identified as a coordinated administrative fleet restart (REGIME_K) rather than an attack, confirmed by a 21:1 ratio in restart ages between returning and non-returning relays.

Significance and Claims

The paper claims to introduce a candidate structural-monitoring framework for behavioral populations that operates without labeled training data, attack signatures, or manual threshold tuning.

Structure Precedes Geometry: The central thesis is validated: the structural organization of the population (the 9-dimensional stiff subspace) is the signal, and geometric metrics are instruments for measuring it.
Detection of Invisible Failure Modes: The framework successfully detected a network failure (connectivity degradation) that did not involve topology changes, which standard monitoring would miss.
Operational Validity: The system achieved a 0.0% FPR on primary detection gates for stable periods and correctly distinguished between population surges, administrative maintenance, and structural fractures using the dual-observer and $\delta_{mg}$ logic.
Limitations: The authors explicitly state they do not claim attack attribution or prediction certainty. The framework detects what changed geometrically, not why. The validation relies on a single externally confirmed event (February 20), which is acknowledged as a constraint of the Tor domain's lack of ground-truth telemetry.

The work reframes network monitoring from "watching for known patterns" to "measuring structural pressure," offering a method to detect anomalies in encrypted, large-scale behavioral populations before they manifest as threshold-crossing events.