Skip to main content
ParalleliX Docs

// Chapter 07 · Protocol

Threat Model

Six ways a node can go bad.

9 min4 sectionsProtocol

Every attack has a named mitigation.

// 7.0 · and we name what we don't defend too

// Threat model · shape

Six misbehaviour categories. Eleven named threats. Three explicit non-defenses.

Detection lives at the coordination layer. Enforcement is uptime-based: a node that fails liveness or validation earns nothing for that period. There is no slashing; staked principal is never seized.

Six misbehaviour categories

// 7.1 · increasing severity · detection at the coordination layer

// Misbehaviour register · 6 categoriesseverity 1 → 5

  1. // 01

    Soft offline

    sev 1/5

    Fails to accept tasks for more than 24h without prior unstake notification

    // detection Coordinator heartbeat monitor

  2. // 02

    Failed PoE

    sev 2/5

    Returns a result whose PoE hash does not match on reconstruction

    // detection Hash-commitment check

  3. // 03

    Wrong result

    sev 3/5

    Returns valid PoE but the output is wrong, caught by redundant re-dispatch

    // detection Redundancy comparison (sampled second node)

  4. // 04

    Provably malicious

    sev 4/5

    Returns deliberately corrupted, manipulated, or fabricated data

    // detection Forensic analysis on returned payload

  5. // 05

    Collusion

    sev 5/5

    Multiple nodes return matching wrong results to bypass the redundancy check

    // detection Statistical analysis of result agreement

  6. // 06

    Sybil identity

    sev 5/5

    Single operator runs multiple node identities to capture more work than allowed

    // detection Behavioural fingerprinting · IP / hardware overlap

Enforcement is uptime-based

// 7.2 · no slashing · the reward model carries enforcement

// 7.2 · From launchAvailable now
  • A non-reconstructing PoE hash is rejected; the request is re-dispatched.
  • Sampled re-dispatch catches a wrong-but-well-formed result and flags the node.
  • Failing liveness or validation earns no uptime credit; that pool share redistributes to online nodes.
  • Persistent failure drops the node from the active set. No slashing; staked principal is never seized.
// 7.2 · As the network opensPlanned
  • Same uptime-based model. Reputation tracking records per-node history as the network opens.
  • Persistent bad actors are flagged and stop earning. Stake stays locked and is returned in full on unstake.

Named threats and mitigations

// 7.4 · eleven vectors · four clusters

// Cluster · Network2 vectors

  • threatEavesdropping on dispatched request payloads

    mitigationEnd-to-end encryption to the recipient node's RSA public key

  • threatImpersonating a node

    mitigationSigned messages verified against the registered public key

// Cluster · Protocol3 vectors

  • threatReturning a forged result

    mitigationHash-commitment check + sampled redundant re-dispatch

  • threatReplay of a previously valid result

    mitigationrequest_id is unique per request; commitment binds result to request

  • threatColluding nodes returning matching wrong results

    mitigationStatistical analysis of agreement; flag both, escalate validation

// Cluster · Operational3 vectors

  • threatOne operator running multiple sybil identities

    mitigationBehavioural fingerprinting, hardware / IP overlap detection

  • threatCoordinator going offline

    mitigationIn-flight tasks pause; documented single point of failure at launch

  • threatCoordinator going dishonest

    mitigationPlanned decentralised validator quorums; at launch, social

// Cluster · Smart contract3 vectors

  • threatCompromise of a tax-destination wallet

    mitigationOwner-key destination rotation (Part VIII.8)

  • threatSmart contract exploit on $PRLX

    mitigationOwner-key transfer pause + audited migration path

  • threatSmart contract exploit on NodeRegistryLocker

    mitigationContract audit + emergency pause; stake and reward balances kept separate so the reward path can never drain principal

What the launch network does not defend

// 7.4 note · three named gaps · honest accounting

// Out of scope · 3 entries · at launch

  • // 01

    Coordinator segmentation honesty

    A dishonest coordinator could segment unfairly. Planned decentralised validator quorums address this.

  • // 02

    Node side-channel attacks

    Out of scope. The node is treated as a trusted-but-verified execution environment.

  • // 03

    Submitted payload correctness

    The network does not validate that a submitted model, simulation, or render scene is semantically correct. It validates that the result matches what the algorithm produces on the payload.

How detection scales

// 7.5 · five mechanisms · O() cost per network size

// Detection cost profiles · 5 mechanisms

  • Hash-commitment verificationScales gracefully

    One SHA-256 reconstruction per request. At 10,000 requests per second the coordinator uses under one CPU core. Essentially free to scale.

    // cost per requestO(1)
  • Sampled redundant re-dispatchScales with throughput

    A fraction of requests run on a second node. A planned calibration target is 5% rate; rate adapts upward when validation disagreements exceed a threshold.

    // cost per requestO(R·t)
  • Statistical agreement (collusion)Does not scale

    Pairwise comparisons across the active node set. A 10,000-node network needs ~50 million comparisons per epoch, so a clustering approximation is planned. A flagged node drops from the active set and earns nothing while its stake sits idle, which backstops this layer.

    // cost per epochO(n²)
  • Behavioural fingerprinting (sybil)Scales gracefully

    Constant-time hardware/IP/network-fingerprint check at registration plus periodic re-check during operation. Scales with the registration rate, not network size.

    // cost per registrationO(r)
  • Reputation trackingScales gracefully

    Incremental per-node counter for uptime and validation success. Memory scales linearly with network size; CPU is negligible. On-chain snapshot at epoch boundaries (planned design).

    // cost memoryO(n)

// Key claim

The mechanisms that scale gracefully defend the most mechanical attacks. The one that does not scale is backstopped by the strongest economic disincentive.

// Where to go next · reading path