CamouflageCheck

PRAXIS: measure, select, confirm, stress-test

An engineering process for choosing a proxy configuration by measurement, then testing how far that choice holds. It does not prescribe one best configuration. It shows how to find one for a given threat model, and how far the evidence for it reaches.

The process

When you test many configurations and report the best one, the data that chose it cannot also be the evidence that it works. PRAXIS keeps selection and confirmation apart.

  1. DiscoverScore many candidate configurations on lab sessions. Phase 13: 8 candidates, 10 sessions each. Phase 15: 3 at 30. Phase 16: 8 at 20.
  2. SelectThe candidate with the lowest PRAXIS score advances.
  3. ConfirmRetest on new sessions that played no part in selection. Phase 17: 150 sessions, 50 per configuration.
  4. Stress-testCheck the winner against holdout detectors, calibrated thresholds, five detector families and five benign references.

Phase 17 compared three configurations. Each contrast changes one component and holds the rest constant.

ConfigurationBrowser profileMultiplexingBrowser parallelismSessionsRole
B baseline_default_offDefaultOff150Baseline
C camouflaged_c2_p1Camouflagedc2150Phase 16 winner
D default_c2_p1Defaultc2150Default-profile control

C against D isolates the browser profile (H1). B against D isolates multiplexing (H2). c2 means proxy multiplexing concurrency of two; parallelism 1 means the browser makes one request at a time.

The score

Four measurements combine into one number. Lower is better.

S = 0.45·detector + 0.35·benign distance + 0.05·overhead + 0.15·max(0, 0.95 − success)
Detector
The detector’s mean confidence, from 0 to 1, that the sessions are proxy traffic.
Benign distance
How far a session’s features sit from public benign traffic. Each feature is log-transformed with its sign kept, standardized by the benign reference’s median and interquartile range, and combined as a straight-line (Euclidean) distance. Distances are then scaled 0 to 1 across the sessions being compared. Lower looks more like ordinary traffic.
Overhead
Capture size in bytes, scaled 0 to 1 across the sessions being compared.
Completion penalty
max(0, 0.95 − success rate). It adds nothing unless fewer than 95% of sessions complete.

The weights are a default engineering policy, not optimized values. The contribution is the multi-criteria structure, and the workbench lets you change the weights and see whether the ranking holds.

Worked example, configuration D in Phase 17: 0.45 × 0.487 + 0.35 × 0.581 + 0.05 × 0.215 + 0.15 × 0 = 0.433.

Features

Every session is reduced to 29 numbers that describe its shape, plus capture size for overhead.

FeatureMeaning
Packets: total, sent, received, direction unknownCounted from the client’s point of view
Payload bytes: sent, received, totalTransport payload, after IP and TCP or UDP headers
Flow durationTime from the session’s first packet to its last
Burst countNumber of runs of consecutive packets in one direction
First 20 signed packet lengthsPayload size of each of the first 20 packets, positive when sent and negative when received

Session duration is included as one total, so strictly the features are payload-independent rather than timing-free. Time between packets is not used.

Detectors and thresholds

The detector used during discovery was the scoring target, so confirmation could not rely on it.

DetectorPhaseRole
Original9Scoring target during discovery
Holdout, baseline-known20Trained on earlier sweep sessions against public benign traffic, with every Phase 17 session excluded
Holdout, strict-unseen20As above, and never saw the selected configuration
Calibrated thresholds20BThresholds set so 0, 0.1, 0.5, 1 or 5% of public benign traffic is flagged
Five detector families30Random forest, extra trees, calibrated linear SVC, gradient boosting, logistic regression

Both holdout detectors are random forests with 500 trees.

Benign references

“Looks like benign traffic” needs a definition of benign traffic. PRAXIS measures it against public captures and checks whether the choice of reference changes the answer.

  • MAWI backbone traces, samplepoint F, 2024-03-01 and 2024-03-02 (10,288 and 10,408 flows). Primary reference, because it is packet-derived like the study’s features.
  • CIC-IDS 2017 benign flows mapped to the same features (2.27 million rows). Secondary check.

Phase 31 rescored every confirmation session against five references: MAWI day 1, MAWI day 2, both days, CIC-IDS 2017, and a balanced MAWI and CIC-IDS mix.

Statistics

  • Mean differences: bootstrap 95% confidence intervals and two-sided permutation tests with 10,000 shuffles
  • Feature contrasts: percent change, pooled standard deviation and standardized effect size
  • Success proportions: Wilson 95% intervals
  • Resolver comparisons: Fisher exact tests with Benjamini-Hochberg adjustment
  • Significance at α = 0.05; statistical evidence and practical size reported separately

34 phases in three cycles

The work ran as a sequence of numbered phases, grouped into three cycles.

Cycle 1, phases 1–10Built the lab and ran the first formal experiments
Cycle 2, phases 11–27Built the selection framework and the software
Cycle 3, phases 28–34Mechanism, robustness and statistics
PhaseWhat it did
9Trained the original detector
12Extracted public benign flow features
13–16Discovery sweeps and candidate scoring
17Confirmation: 150 new sessions
20, 20BHoldout detectors and calibrated thresholds
21Resolution-path test: 90 sessions
24, 25, 27API service, PCAP analyzer and dashboard
29, 29BMechanism analysis and artifact control
30Five detector families
31Five benign references
34Inferential statistics

The results reported on this site come from cycles 2 and 3.

Software

PRAXIS is implemented as a shared Python library, a FastAPI service, a Streamlit dashboard and a PCAP analyzer, all in the public repository. The workbench on this site reimplements feature extraction and scoring in the browser so anyone can use them without installing anything. With the study’s 150 confirmation sessions it reproduces the Phase 31 scores to four decimal places; see Reproduce.