CamouflageCheck

What the study found

A configuration’s effect depends on where, and against what, you measure it. One setting cut detector confidence substantially, yet independent detectors still flagged every session.

Summary

QuestionComparisonResultOutcome
H1, browser profileCamouflaged against default, both multiplexing c2Camouflaged scored higher, not lower, under all 7 detectorsNot supported
H2, multiplexingMultiplexing c2 against offConfidence fell by 0.29 to 0.51; detection stayed at 100% at every calibrated thresholdSupported for confidence, not for classification
H3, resolution pathAlternate mapping against standard DNS30 of 30 in bothNo difference: compatible

The claim, exactly: a reduction in detector confidence under a controlled lab threat model. Not detector bypass, and not evidence of real-world censorship resistance.

H1: browser profile

Phase 17 compared the camouflaged and default browser profiles with multiplexing fixed at c2 and parallelism 1, 50 sessions each.

Original detector+0.001Holdout, strict-unseen+0.012Holdout, baseline-known+0.002Random forest+0.026Extra trees+0.002Linear SVC, calibrated+0.010Logistic regression+0.001
Camouflaged minus default, mean detector probability. Gradient boosting scored both profiles identically and is not shown.

No detector scored the camouflaged profile lower. The traffic shows why: both produced a mean duration of 0.188 seconds and about 54 packets. The camouflaged profile also had higher overhead, and its PRAXIS score was 0.030 worse. Both completed every session.

H1 not supported Browser camouflage acts at the browser layer, and a flow detector does not see that layer. The archived results do not include a direct inferential test of this contrast, so the conclusion is “no demonstrated advantage,” not proof of zero effect.

H2: multiplexing

Phase 17 compared multiplexing off (B) with multiplexing c2 (D), default profile and parallelism 1, 50 sessions each. This comparison carries the most information in the study.

Original (selection) detector1.0000.487Holdout, strict-unseen0.9600.666Holdout, baseline-known0.9980.663Random forest0.9620.580Extra trees0.9950.839Linear SVC, calibrated0.9570.947Gradient boosting1.0001.000Logistic regression0.9950.996
Multiplexing offMultiplexing c2
Mean detector probability by detector.

Under the original detector, mean probability fell from 1.000 to 0.487, a reduction of 0.513 (p < 0.001). It also fell under both holdout detectors, by 0.294 and 0.335. Every session still completed, and overhead dropped.

The size of the drop depends on the detector family: 0.38 for random forest, 0.16 for extra trees, and essentially zero for linear SVC, gradient boosting and logistic regression. Tree models split on count thresholds, so a 20% shift crosses split points. Linear models add up correlated features and barely move.

Lower confidence did not change the decision

DetectorMean probabilityThresholdSessions flagged (of 50)
Original (selection) detector0.4870.53 (6%)
Holdout, strict-unseen0.666Calibrated to 0–5% false positives50 (100%)
Holdout, baseline-known0.663Calibrated to 0–5% false positives50 (100%)
Random forest0.5800.549 (98%)
Extra trees0.8390.550 (100%)
Linear SVC, calibrated0.9470.550 (100%)
Gradient boosting1.0000.550 (100%)
Logistic regression0.9960.550 (100%)

The lowest holdout probability of any multiplexed session was 0.58, well above the strictest calibrated thresholds of 0.216 and 0.275. The detector used to pick the winner was fooled most of the time; independent detectors, set the way an operator would set them, caught everything. That gap is why the confirmation step exists.

H2 supported for confidence, not for classification The result holds for c2 multiplexing at parallelism 1. Earlier static multiplexing tests at c2, c4 and c8 showed no change in identifiability.

What changed in the traffic

FeatureMultiplexing offMultiplexing c2Change
Duration (s)1.1430.188−83.5%
Packets68.853.8−21.9%
Bursts43.133.8−21.4%
Payload bytes15,34915,548+1.3%
Capture size (bytes)21,49320,321−5.5%

Multiplexed sessions were compressed: much shorter, with about a fifth fewer packets and bursts, carrying the same payload. The traffic became a different shape, not an invisible one. In the Phase 29B artifact control, a detector trained only to tell the two configurations apart separated them perfectly (AUC 1.00), even with duration and packet counts removed and even from the first 20 signed lengths alone.

H3: resolution path

Phase 21 held the selected configuration’s transport constant and ran 30 curl sessions in each of three resolver modes, all to one internal hostname. Port 53 was blocked for the blocked mode; the alternate mapping came from a local hosts-file entry.

Standard lab DNS30 of 30DNS blocked (port 53), no fallback0 of 30Alternate hostname mapping30 of 30
Session success by resolver mode.

Standard DNS and the alternate mapping both succeeded in 30 of 30 sessions: risk difference 0, Fisher p = 1.000. The blocked mode failed in all 30, which confirms the block worked. The Wilson 95% interval for each working mode is 0.886 to 1.000.

H3: no difference Swapping the resolution path did not degrade the configuration. No difference is not proof of equivalence: a drop of up to about 11 points cannot be ruled out. This is a compatibility test with a hosts-file mapping, not evidence of censorship resistance, and detectability under the alternate path was not measured.

“Looks benign” depends on the reference

Phase 31 rescored the confirmation sessions against five benign references. The selected configuration won under all five, but whether it moved toward or away from benign traffic depended on which reference was used.

Benign referenceDistance change, D against BScore, DLowest score
MAWI day 1+0.1100.433D
MAWI day 2+0.1110.433D
MAWI, both days+0.1050.432D
CIC-IDS 2017, mapped−0.1200.372D
MAWI and CIC-IDS, balanced+0.0120.416D

Against packet-derived MAWI traces, the selected configuration moved away from benign traffic; against mapped CIC-IDS 2017, it moved toward it. Benign similarity is a relation to a reference, not a property of the traffic. D won on detector confidence and overhead, not on similarity.

Conclusions

  1. Layer matters. Browser-profile camouflage did not lower network-level detector confidence under any of 7 detectors.
  2. Configuration matters. Static multiplexing changed nothing; c2 at parallelism 1 cut confidence by 0.29 to 0.51, depending on the detector.
  3. Confidence is not classification. Detection stayed at 100% at every threshold calibrated to public false-positive rates.
  4. Evaluation choices matter. Effects depend on detector family and benign reference; single-model results can mislead.
  5. The resolution path is compatible. The selected transport stayed fully operational over an alternate mapping, 30 of 30 against 30 of 30.

All five hold within the passive, non-timing, laboratory threat model. A circumvention defense offers meaningful protection only when its advantage persists across aligned detection layers, detector families, operating thresholds and representative reference traffic.

Limitations

LimitationWhat it limits
Training confound: proxy sessions came from the lab, benign negatives from public captures100% detection may partly reflect the capture environment rather than proxy use. The B against D comparison is less affected, since both came from the same lab.
Near-replicate sessions: scripted workload, fixed destinations, duration spread of 0.003 sp-values and effect sizes show separation, not operational effect size
Static, shape-only observerNo claim against timing-aware, handshake-aware or adaptive censors
Confirmation design: parallelism fixed at 1; randomized run order not documentedH2 confirmed for c2 at parallelism 1 only; drift not ruled out
Narrow resolver test: one hostname, hosts-file mapping, one block typeShows compatibility only, not censorship resistance or detectability

Next steps

  1. Same-environment negatives. Train detectors on benign traffic captured in the same lab, so they cannot lean on environment differences.
  2. Realistic variability. Variable workloads and many destinations instead of near-replicate sessions.
  3. Factorial transport design. Confirm multiplexing and parallelism together, not at parallelism 1 only.
  4. Adaptive adversaries. Censors that retrain on the configurations they observe (Abolfathi et al., 2024).
  5. Real decentralized naming. Production systems under poisoning, NXDOMAIN injection and null routing, measuring availability, latency and detectability.