obscura/research/uncontrollable/proxy-detectionobscura/research/uncontrollable/proxy-detection

Category: Cross-Layer / Meta-Detection

---

1. Description

The act of protecting oneself creates detectable patterns. Privacy tools are paradoxical: the more aggressively they modify signals, the more they create new signals that identify "privacy tool user." This document analyzes how fingerprinting services detect Obscura-like proxies through inconsistencies and anomalies.

---

2. Detection Vectors

2.1 Cross-Layer Profile Inconsistency

The most powerful detection: comparing signals across layers.

Entropy: Very High can definitively detect proxies

2.2 TLS Proxy Detection

2.3 JavaScript Override Detection

// Detect if navigator properties have been overridden
const uaDescriptor = Object.getOwnPropertyDescriptor(
  Object.getPrototypeOf(navigator),
  'userAgent'
)
if (uaDescriptor && uaDescriptor.get !== originalGetter) {
  // Override detected!
}

// Alternative: compare multiple signal sources
// navigator.userAgent vs navigator.userAgentData
// If they disagree, one has been overridden

Additional techniques:

2.4 Feature Blocking Detection

Pattern analysis: A user blocking WebGL + AudioContext + WebRTC + Canvas is almost certainly using a privacy tool.

2.5 Behavioral Anomalies

2.6 IP-Based Detection

2.7 DNS-Based Detection

---

3. Statistical Detection

Sophisticated attackers use machine learning to detect anomalies:

# Conceptual ML detection model
features = [
    'tls_fingerprint_mismatch',
    'feature_blocking_count',
    'js_override_detected',
    'header_consistency_score',
    'network_anomaly_score',
    'behavioral_anomaly_score',
    'ip_reputation_score',
]

model = RandomForestClassifier(n_estimators=100)
# Trained on labeled dataset of real users vs privacy tool users
prediction = model.predict(features)
# Probability that this traffic is behind a privacy tool

The ML model can detect subtle correlations humans would miss e.g., "users who block WebGL and use Cloudflare DNS and have consistent headers are 94% likely to be using a privacy tool."

---

4. Attacker's Strengths

---

5. Attacker's Weaknesses

---

6. Mitigations for Obscura

6.1 What Obscura Can Do

6.2 What Obscura Cannot Do

6.3 Recommended Approach

1. Profile consistency validation       (highest ROI  prevents most detection)
2. Real CA certificates                  (avoid self-signed CA detection)
3. Automatic fingerprint updates         (keep up with browser changes)
4. Strategic signal leaking              (make profile look more natural)
5. Accept  some detection is inevitable (document residual risk)

---

7. The Meta Conclusion

The uncomfortable truth: a perfect proxy-based anti-fingerprinting system is impossible because the existence of the proxy is itself a signal. Every protocol modification, every injected script, every blocked API creates a detectable pattern.

This does not mean Obscura is useless. It means:

1. Obscura protects against opportunistic fingerprinting the vast majority of tracking scripts
2. Obscura does not protect against determined fingerprinting sophisticated services with ML detection
3. Obscura + Tor Browser provides defense in depth one for casual browsing, the other for sensitive sessions

The goal is not invisibility. It is raising the cost of identification to the point where most trackers give up.

---

8. Research References

• Arshad, S. et al. (2016). "A Measurement Study of Web Proxy Detection." IEEE S&P 2016.
• Englehardt, S. & Narayanan, A. (2016). "Online Tracking: A 1-million-site Measurement and Analysis." CCS 2016.
• Vastel, A. et al. (2018). "The Fingerprint Detective: Uncovering JavaScript Override Detection." PETS 2018.
• Laperdrix, P. et al. (2020). "Browser Fingerprinting: A Survey." ACM TOIT.
• Tschantz, M. et al. (2016). "On the Comprehensibility of Browser Fingerprinting." WEIS 2016.