USENIX Security ’26 · Peer-reviewed research

Tor hides where you go.
What can the rhythm of packets still reveal?

Layered encryption conceals message payloads and obscures destination routing. Yet a passive observer monitoring the client’s local connection can still record burst lengths, packet directions, and inter-arrival timing. A newly peer-reviewed framework cuts the operational cost of maintaining a surveillance watchlist: onboarding an unseen website requires only 5 to 20 fresh visits, even when the user browses with multiple tabs simultaneously.

5–20
traces needed to onboard a new monitored site
down from hundreds or thousands in older models
73.6%
precision across 5 simultaneous tabs with just 5 shots
under controlled multi-tab evaluation
47.0%
precision after 2 months of real web drift with 5 tabs
dropping below a coin toss

What the attacker can and cannot observe

To assess website fingerprinting realistically, one must begin with the observer’s position. In Tor’s threat model, a client routes encrypted traffic through a three-relay circuit. The attacker evaluated in the paper is strictly passive: positioned on the client-side link, such as a local Wi-Fi gateway, an ISP choke point, or the wire directly preceding the Tor entry guard.

This observer never decrypts payloads, injects packets, modifies data, or introduces artificial delays. All Tor traffic is packaged into fixed 512-byte cells. Nevertheless, as a browser renders a modern webpage, it requests resources—HTML, stylesheets, images, scripts—in characteristic sequences, producing a distinctive temporal fingerprint of inbound and outbound bursts.

Page Payload & Content
Fully end-to-end encrypted across three relays; opaque to the observer.
Destination Server & IP
Obscured by Tor onion routing; the observer sees only the first guard relay.
Direct User Identity
Not provided by packet metadata; requires fusion with external identifiers.
👁 VISIBLE
Individual Packet Directions
Inbound (-1) vs outbound (+1) packet sequences logged per millisecond.
👁 VISIBLE
Burst Cell Counts
Number of 512-byte cells transmitted before the next directional pause.
👁 VISIBLE
Inter-Arrival Timings
Precise millisecond gaps between client requests and server responses.

The target watchlist maintenance bottleneck

For years, practical website fingerprinting hit a steep logistical wall: target maintenance. Real websites continually change layouts, advertisements, and third-party assets. Moreover, actual users rarely browse in artificial single-tab isolation; they routinely keep three, four, or five tabs open at the same time.

Conventional fingerprinting classifiers attempted to recognize the entire trace as a massive combinatoric mixture. Whenever an intelligence actor wanted to monitor a newly emerged whistleblower portal or exile media outlet, it had to harvest thousands of fresh traffic traces and retrain heavy neural networks from scratch.

The MMF architecture reframes the problem entirely: instead of predicting the full combinatoric label, it asks independently whether features matching a monitored target are present inside the mixture. By applying metric-based few-shot learning, the observer can onboard a newly interesting site using as few as 5 to 20 recorded visits.

CLASSICAL

Conventional Combinatoric Fingerprinting

  • Assumes 1 trace = 1 website or relies on rigid combination sets.
  • Requires hundreds or thousands of fresh traces per new website.
  • Demands full model retraining whenever the target watchlist updates.
  • Collapses rapidly when users open an unknown number of concurrent tabs.
NEW MMF

MMF Framework (USENIX Security ’26)

  • Treats multi-tab browsing as a composite mixture of independent signals.
  • Onboards newly monitored websites with only 5 to 20 recorded examples.
  • Bypasses heavy retraining by comparing metric feature embeddings.
  • Operates robustly even as the user’s tab count fluctuates in real time.

The Packet Oscilloscope: multi-tab interleaving and detection

Explore firsthand how simultaneous browsing tabs interleave encrypted bursts, and how the quantity of training traces (shots) impacts the classifier’s ability to detect a watched target.

● ACTIVE BURST OSCILLOSCOPE TOR CELL STREAM · 512B CELLS
Packets originating from monitored target
Packets from other open tabs (background noise)
Outbound packet (upload)
Inbound packet (download)
Target presence confidence score: 62.4%
Total 512-byte cells in inspection window: --
Classifier output verdict:
● Target Detected

Comparative performance in open-world settings

The authors evaluated MMF against leading academic baselines in both controlled conditions and realistic open-world environments containing thousands of unmonitored background sites.

Precision at k (P@k) in 4-tab open-world testing under distribution shift

MMF (New Framework) 62.36%
3.51× higher precision than the closest baseline
BAPM (2024 Baseline) 17.75%
Deteriorates sharply in multi-tab conditions
TMWF (Baseline) 15.20%
Struggles with interleaved packet bursts
ARES (Baseline) 13.43%
Low precision in multi-label decomposition
FMWF (Baseline) 6.20%
Near-total loss of signal in open-world testing

Why 47% precision stops the 'superpower' narrative

The most instructive finding in the paper is not peak lab accuracy, but the measurement of real-world drift. To replicate authentic surveillance constraints, the researchers collected traffic over approximately two months, allowing websites, CDNs, and network routes to shift naturally.

Even when provided with 20 fresh traces per new website, MMF’s precision on the drift dataset dropped to 59.1% for 3 tabs, 54.5% for 4 tabs, and 47.0% for 5 tabs. When a user has 5 tabs active, the probability of a correct passive detection falls below a coin toss.

This boundary demonstrates the inherent limits of passive traffic analysis: it is not omniscient, and it rapidly decays without continuous trace collection. The strategic shift is strictly economic: the operational burden of keeping a watchlist current dropped from thousands of traces to a few dozen.

3 concurrent tabs
59.15%
Moderate precision; burst interleaving is partially disentangled.
4 concurrent tabs
54.50%
Substantial degradation; nearly half of all inferences are ambiguous.
5 concurrent tabs
47.02%
Below 50%; complex traffic interleaving successfully obscures the target signature.

Institutions, attribution, and capability signals

The paper is co-authored by researchers from China’s National University of Defense Technology (NUDT) and Tsinghua University. NUDT’s institutional structure places it directly subordinate to China’s Central Military Commission.

This institutional affiliation serves as a clear capability signal regarding high-level state interest in traffic analysis and anonymity-network research. However, it does not demonstrate that this specific framework is currently deployed operationally by Chinese military or intelligence agencies.

The authors deposited their source code and reproduction benchmarks openly on Zenodo in May 2026 and framed the research within privacy and defensive security boundaries, explicitly noting the risks of surveillance misuse against journalists and dissidents.

🛡 Crucial Distinction: Detection vs Identification

This framework detects statistical consistency with a target website; it does not directly identify an individual human being. Associating a detected traffic burst with a named person requires correlating the observation with subscriber records, gateway IP logs, device telemetry, or physical intelligence.