The Processed Replay Blindspot: Why Modern Bot Defenses Fail

1. The Evolution of Bot Detection

For over two decades, behavioral biometrics in web applications has centered on a single fundamental premise: machine-generated automation lacks the physical subtleties, neuromuscular tremor, and motor hesitation of a human hand.

Defenders developed three main architectural layers:

1. Behavioral Machine Learning (e.g. Random Forest / SVM): Detects statistical regularities such as over-smooth Bézier curves, constant velocity, and unrealistic acceleration derivatives (jerk variance).
2. Historical Similarity / Embedding Comparison (e.g. Dynamic Time Warping, LSTM Autoencoders): Flags exact duplicates or near-identical replayed sessions against a stored trajectory library.
3. Context & Click-Target Validation: Checks whether click coordinates align with legitimate, interactive UI bounding boxes in a logical sequence.

[!NOTE] While these three layers successfully eliminate 99% of basic automation (Puppeteer, Selenium, raw HTTP replay), they suffer from a fatal theoretical assumption: that the adversary generates artificial motion or repeats exact previous trajectories.

---

2. The Unsolved Gap: Processed Replay

Two independent IEEE studies exposed the exact boundary where traditional defenses collapse:

• Salman & Bicakci (IEEE Access, 2026): Evaluated a 6-level ladder of bot sophistication across 1,140 real sessions. Every single-channel defense succeeded against synthetic bots-until processed replay ($A_6$), where an adversary splices fragments from multiple genuine human sessions. Salman & Bicakci declared processed replay explicitly out of scope for existing biometrics.
• Sadeghpour & Vlajic (IEEE CSR, 2024 - ReBotDetector): Built an LSTM-autoencoder for session replay, but acknowledged it only detects replayed sessions with near 1.0 cosine similarity to a single recorded session. It cannot detect composite sessions assembled from multiple human donors.

4×4 Multi-Detector Benchmark Matrix

The table below demonstrates how traditional detectors perform across the attack sophistication ladder compared to CORROBORATE:

[!WARNING] Because every individual trajectory fragment in an $A_6$ attack was originally generated by a real human hand, single-channel mouse statistics pass all statistical normality tests.

---

3. How Adversaries Construct CHIMERA Sessions

An adversary executes a 5-step pipeline to synthesize a deceptive session:

1. Pause-Boundary Segmentation: Cuts donor sessions at stationary dwell points ($v \approx 0$) so jerk and acceleration derivatives remain smooth.
2. Multi-Source Splicing: Merges fragments from 3+ distinct human donors to defeat DTW similarity matching.
3. Similarity Transform Retargeting: Translates and rotates trajectories to target UI coordinates.
4. Independent Typing Overlay: Injects foreign keystrokes onto the spliced mouse timeline.

jsonCopy
{
  "event_type": "pointermove",
  "timestamp": 1420.5,
  "x": 412,
  "y": 198,
  "velocity": 0.04,
  "source_donor": "donor_02_mouse",
  "concurrent_events": [
    {
      "event_type": "keydown",
      "key_code": 65,
      "timestamp": 1421.0,
      "source_donor": "donor_07_keys"
    }
  ],
  "anomaly": "Parallel Pointer-Key Collision (PPX violation)"
}

---

4. The CORROBORATE Paradigm Shift

CORROBORATE solves this by shifting focus from intra-channel shape to cross-modal temporal coupling (CMCC):

• Attentional Exclusivity: Humans have a single cognitive processor. When typing, cursor motion halts ($ ext{MIDT} \approx 1.0$).
• Homing Operator Bound: Moving the dominant hand between keyboard and mouse incurs a mandatory physical cost ($Delta t \approx 400\text{ms}$).
• Causal Ordering: Visual targeting precedes cursor movement, which precedes element focus ($ ext{PPF} > 0$).

javascriptCopy
// Cross-Modal Coherence Evaluation Loop
function evaluateWindowCoherence(eventsWindow) {
  const midt = calculateMouseIdleRatio(eventsWindow);
  const coc = evaluateHomingLatency(eventsWindow);
  const ppx = computeConcurrentCollisions(eventsWindow);

  // Composite window score
  const windowScore = (0.4 * midt) + (0.35 * coc) + (0.25 * (1 - ppx));

  if (windowScore < CALIBRATED_THRESHOLD) {
    return { verdict: "FLAGGED_SEAM", confidence: 1 - windowScore };
  }
  return { verdict: "AUTHENTIC", confidence: windowScore };
}

Key Takeaways

• Single-channel bot detection is fundamentally vulnerable to spliced human replay.
• Cross-modal coupling provides mathematical proof of single authorship without tracking PII.
• Zero user disruption: Verification occurs passively in the browser in <20ms.