Risk Detection
The Cencera Trust Protocol combines four independent signal domains into a single probabilistic trust score. Gaming all four simultaneously is exponentially harder than gaming any one in isolation.
Design Principles
Trust as a First-Class Primitive
Not a feature or binary flag, but a continuously computed probabilistic score with full structured metadata.
Graceful Degradation
Even with limited data, a labeled partial score is more useful than null. Nascent entities receive clearly labeled limited scores.
Tiered Explainability
Category-level signal triggers published publicly; signal-level detail for enterprise partners only — transparency without enabling adversarial gaming.
Continuous Learning
Every confirmed exploit triggers retroactive rescoring of all similar entities within 4 hours.
Platform Autonomy
CENCERA scores; consuming platforms enforce. Each partner configures its own thresholds, signal weights, and lookback periods.
Cross-Chain Architecture
Scores are chain-scoped by default. Negative signals propagate more strongly across chain links — attackers cannot reset reputation by switching chains.
Why Existing Approaches Fall Short
| Approach | Mechanism | Core Limitation |
|---|---|---|
| Static Blacklists | Block known malicious addresses | Cannot detect novel attacks; attackers rotate addresses trivially |
| Transaction Simulation | Simulate a transaction before execution | Blind to slow-burn exploits; no behavioral history; single-chain |
| Binary Screening | Pass/fail flag on a transaction | No probabilistic nuance; no trajectory; no cross-chain awareness |
The Four Signal Domains
On-Chain Signals — Cryptographically Verified
- •Transaction frequency and value patterns over time
- •Contract interaction graphs using graph-based neighbor risk propagation
- •Unlimited approval grants and revocation behavior
- •Liquidity provisioning and withdrawal timing relative to price movements
- •Historical adjacency to known exploit contracts
- •Contract upgrade and proxy change events
- •Governance participation patterns
Market Signals
- •Liquidity depth, volatility, and depth at various price impact levels
- •Token distribution concentration — Gini coefficient, top-holder percentage
- •Slippage anomalies and sandwich attack frequency
- •Sudden supply changes or fee/tax modifications
- •Price-to-liquidity ratio as a manipulation signal
Off-Chain Signals — Social & Identity Context
- •Website domain age, registration anonymity, and DNS consistency
- •Social account authenticity — follower quality, engagement, cross-platform consistency
- •Developer identity continuity across projects
- •GitHub repository history — commit authenticity and sudden deletions
- •Audit status from recognized security firms
AI-Derived Signals
- •Bytecode similarity clustering — detects obfuscated exploit variants that evade exact-match detection
- •Behavioral anomaly scoring — deviation from an entity's own historical baseline
- •Exploit pattern recognition — reentrancy setups, flash loan sequences, governance manipulation matched against live streams
- •Temporal risk trend analysis — monotonically increasing risk interpreted as attack staging
- •Novel Behavior Flag — unknown patterns routed to human review with provisional score
Adversarial Resilience
Multi-Signal Correlation
Each domain is gameable in isolation; all four simultaneously is exponentially harder.
API Probing Prevention
Tiered explainability + rate limiter escalation prevents systematic reverse-engineering.
Data Source Poisoning Guard
Multi-provider cross-validation; write-protected exploit library requiring multi-party authorization.
Sybil Detection
Star-topology interaction graphs (bot clusters) are flagged immediately on detection.
Behavioral Mimicry Detection
Attack staging patterns detected before execution completes via temporal risk analysis.
Retroactive Rescoring
Confirmed exploit triggers retroactive rescore of all similar entities within 4 hours + webhook notifications.
Prior Approaches & Limitations
A survey of existing blockchain security systems reveals a consistent pattern: each optimizes for one signal type while remaining structurally blind to others. CENCERA is designed to address all of these failure modes simultaneously.
Address Blacklisting
Static address blacklists operate on exact-match: an address either appears in the registry or it does not. An address can only be blacklisted after harm is confirmed — every novel attack succeeds at least once. Address rotation is trivially inexpensive, allowing adversaries to deploy fresh addresses with no accumulated history.
Transaction Simulation
Simulation intercepts pending transactions and runs them in a sandbox. It can detect exploits that produce observable state changes within a single transaction — but cannot evaluate behavioral history over time, detect slow-acting exploits across multiple blocks, or assess entities that haven't yet interacted with the user.
Rule-Based Anomaly Detection
Rule-based systems encode expert knowledge as deterministic rules (flag unlimited approvals, flag concentrated liquidity, etc.). Because rules are public or reverse-engineerable, a motivated adversary can construct an attack satisfying each individual rule while still executing harmful behavior. Rules also evaluate signals in isolation with no cross-signal reasoning.
Single-Chain Identity Systems
Soulbound token or attestation-based identity systems create persistent non-transferable credentials — but reputation on Ethereum carries no formal weight on BNB Chain or Arbitrum. An adversary can build a clean reputation on one chain while staging an attack on another, with no mechanism for risk to propagate across the chain boundary.
Graph-Based Risk Propagation
Graph methods propagate risk from known-bad nodes through transaction graphs to neighbors. They represent a meaningful advance over flat blacklisting but remain reactive: they require at least one confirmed bad node as a seed, and their effectiveness is bounded by transaction graph density. Entities that stage attacks through intermediary contracts evade graph detection.
The Cencera Trust Protocol is designed around the observation that no single signal domain is sufficient. Trust must be treated as a longitudinal, multi-signal, probabilistic quantity rather than a binary attribute or a point-in-time assessment. The four signal domains address each failure mode above: on-chain behavioral history addresses temporal exploitation, market signals address liquidity manipulation, off-chain signals address identity continuity, and AI-derived signals address adversarial adaptation through bytecode similarity analysis and behavioral baseline modeling.