Why traffic shape gives a proxy away
In some countries, internet censorship cuts people off from information and from each other. Encrypted proxies get around the filters, but encryption hides what a connection carries, not what it looks like.
Where the problem comes from
People in censored networks use VPNs and encrypted proxies to reach blocked sites and services. Using one can itself make them a target, so the proxy has to work and also avoid standing out.
A network operator does not need to decrypt a connection to recognize it. Packet counts, sizes, directions and bursts are visible to anyone on the path, and a classifier trained on them can pick out a circumvention connection and block it.
So the engineering question is whether changing a proxy’s configuration makes a connection harder to detect without breaking the user’s ability to communicate. In practice, engineers often assume that switching on a camouflage or obfuscation feature provides protection. This research replaces that assumption with measurement.
Encryption hides content, not shape
Flow-level classifiers already detect obfuscated proxies with high accuracy.
| Study | What was detected | Reported result |
|---|---|---|
| AlOmar et al., 2025 | Obfuscated Tor transports (obfs4, Snowflake), under class imbalance | Above 98% accuracy |
| Ma et al., 2024 | Encrypted-proxy traffic, by passive flow-context analysis | 98.8% detection at under 0.2% false positives |
| He and Li, 2025 | Shadowsocks, from packet-size and timing distributions | Above 98% F1 |
The closest work is Xue and colleagues at USENIX Security 2024, who fingerprinted obfuscated proxies through the TLS handshake carried inside the tunnel. Padding and stream multiplexing changed traffic patterns but left markers such as round-trip counts and initial burst sizes. If multiplexing alone does not remove identifiability, does its effect depend on how it is configured? That is the question behind the second research question.
The evaluation gap
Each of these studies uses its own data, features, classifier and metrics. A claim that a configuration reduces detectability therefore cannot be checked against another study (Jeřábek et al., 2024; Papadogiannaki and Ioannidis, 2021). Jeřábek and colleagues showed that published DNS-over-HTTPS detectors report high precision but cannot be compared directly, because their datasets and methods differ.
- Problem
- There is no standardized empirical basis for deciding whether browser- and transport-level configuration changes reduce non-timing proxy detectability while sessions keep working.
- Thesis
- Circumvention configurations should be selected and validated empirically, across detector confidence, classification, compliance and robustness, rather than assumed effective because a mechanism is switched on.
- What is new
- Not a new camouflage mechanism. PRAXIS is an integrated process that scores candidate configurations on several objectives, separates discovery from out-of-sample confirmation, and stress-tests the winner across detectors, thresholds and benign reference data.
Three levers, one observer
A circumvention client controls three things. Each acts at a different layer, but the observer sees only one layer: the shape of packets and flows. For each lever, the real question is whether a change at its layer reaches what the observer sees.
Browser-profile camouflage is studied at the browser layer. Kalantari and colleagues’ Browser Polygraph (IMC 2024) found that extension-based camouflage lowers fingerprint entropy only partly and can add variability of its own. Whether that camouflage changes anything a network detector sees had not been tested. For the resolution path, Jeon and Park (2026) compare Namecoin, Handshake, Blockstack and ENS and note their dependence on gateways and non-native resolution paths, but say little about whether a traffic configuration keeps working when the resolution path changes.
Threat model
The observer is passive and sits on the path between client and proxy. It cannot decrypt payloads and, in this study, does not use timing between packets. It applies a fixed, trained classifier and flags sessions at thresholds set by how many false positives it can tolerate.
What the observer uses
- Packet counts and direction
- Payload bytes and burst count
- First 20 signed packet lengths
- Session duration, as one total
- Thresholds set by tolerable false positives
What it does not use
- Decrypted payloads
- Time between packets
- Handshake contents, such as cipher suites and extensions
- Retraining against the configurations it sees
Two assumptions matter for the results: the adversary is static and does not retrain, and it sees traffic shape only.
Research questions and hypotheses
- RQ1. To what extent do browser profile and proxy configuration choices reduce non-timing proxy detection confidence while preserving session compliance?
- RQ2. How do multiplexing mode and browser-side parallelism affect proxy traffic identifiability under non-timing traffic analysis?
- RQ3. Can the best traffic camouflage configuration be integrated with an alternate resolution path to preserve access and compliance under DNS blocking?
| Hypothesis | Manipulated | Measured | |
|---|---|---|---|
| H1 | Browser profile configuration reduces non-timing proxy detection confidence when proxy configuration is held constant. | Browser profile, default or camouflaged; multiplexing fixed at c2, parallelism 1 | Detector probability |
| H2 | Multiplexing mode and browser-side parallelism reduce non-timing proxy detection confidence. | Multiplexing off or c2; default profile, parallelism 1 | Detector probability, detection rate |
| H3 | Session success differs between standard lab DNS and the alternate resolution path when the selected configuration is held constant. | Resolver mode: standard, blocked, alternate mapping | Session success |
H1 and H2 test confidence, a probability from 0 to 1. Classification, the yes-or-no decision at a threshold, is examined separately.
Scope
Everything was measured in a controlled Amazon EC2 lab with machine-generated sessions. One session is one observation.
In scope
- Browser profile: default or camouflaged
- Proxy multiplexing and browser-side parallelism
- Resolver path: standard DNS, blocked, alternate mapping
- A passive observer using packet and flow features without inter-packet timing
Out of scope
- Live state-level or adaptive censors
- Timing-based and handshake-content classifiers
- Production decentralized naming (ENS, Namecoin, Handshake)
- DNS poisoning, NXDOMAIN injection, null routing
- Human subjects and user behavior
Key literature
- Ma, X., Qu, J., Shi, M., et al. (2024). Website fingerprinting on encrypted proxies: A flow-context-aware approach and countermeasures. IEEE/ACM Transactions on Networking, 32(3), 1904–1919.
- Xue, D., Kallitsis, M., Houmansadr, A., and Ensafi, R. (2024). Fingerprinting obfuscated proxy traffic with encapsulated TLS handshakes. 33rd USENIX Security Symposium, 2689–2706.
- Kalantari, F., Zaeifi, M., Safaei, Y., et al. (2024). Browser Polygraph: Efficient deployment of coarse-grained browser fingerprints for web-scale detection of fraud browsers. Proceedings of the ACM Internet Measurement Conference 2024, 681–703.
- Jeřábek, K., Hynek, K., and Ryšavý, O. (2024). Comparative analysis of DNS over HTTPS detectors. Computer Networks, 247, 110452.
- Jeon, J., and Park, S. (2026). A survey on traditional DNS and blockchain-based DNS: Comparative analysis, challenges, and future directions. Applied Sciences, 16(2).
Other works cited on this site (AlOmar et al., 2025; He and Li, 2025; Papadogiannaki and Ioannidis, 2021) are referenced in full in the praxis.