Single-modality detection degrades across models
SONAR Benchmark, ACM 2024-2025
OpenAI's TTS is detected only 78% of the time by the best available classifier. Models trained on older datasets lose up to 43% performance against newer TTS systems.
// HOW IT WORKSFrom challenge
From challenge
to on-chain proof.
Capture, feature extraction, and proof generation all run on your device. The validation path receives a statistical summary and transient challenge evidence. Solana stores protocol state, commitments, and proofs.
// VERIFICATION FLOW
12 seconds. The rest is automatic.
Speak a phrase, trace a curve. Feature extraction, proof generation, and on-chain verification run automatically.
01—Challenge
01—Challenge
A random word phrase and Lissajous curve generated fresh for each session. No two sessions share the same challenge.
Each session generates a unique 5-word phrase from a curated English vocabulary and a unique Lissajous curve from random mathematical parameters. The user speaks the phrase while tracing the curve for 12 seconds. The challenge elicits natural behavioral data (voice prosody, hand tremor, touch pressure) rather than testing memory or speed.
02—Capture
02—Capture
Three sensor streams record in parallel: voice, touch, and motion. 12 seconds of simultaneous behavioral data.
The Pulse SDK accesses the device microphone, accelerometer, gyroscope, and touch digitizer. All sensors record in parallel for 12 seconds. Raw motion and touch recordings stay in device memory and are destroyed after feature extraction. Derived statistical summaries leave the device for validation, along with the spoken-phrase audio, which the server transcribes and discards immediately. On desktop, motion sensors are unavailable. Mouse pointer dynamics provide equivalent kinematic features.
03—Extract + Score
03—Extract + Score
308 statistical features across voice, motion, and touch. Anatomical signal (formants, MFCCs, voice quality) alongside anti-synthesis traces (jitter, shimmer, HNR).
Audio contributes 170 features, including F0 statistics, MFCCs, LPC coefficients, formant trajectories, and voice quality. Motion contributes 81 features, including jerk, jounce, band energies, tremor peaks, and cross-axis covariance. Touch contributes 57 features, including velocity, pressure derivatives, curvature, and path efficiency. The private service evaluates these values under the current policy. On desktop, mouse dynamics fill the kinematic feature shape.
04 - Hash
04 - Hash
SimHash projects the feature summary into a 256-bit fingerprint for continuity research.
SimHash projects the expanded feature vector across fixed hyperplanes. The result is a comparable 256-bit fingerprint. Current research measures whether same-person captures stay close enough while different people and synthetic inputs separate reliably.
05—Commit
05—Commit
Poseidon(fingerprint || salt) produces the TBH commitment. The fingerprint and salt stay on-device.
The SDK generates a large random salt. Poseidon commits to the fingerprint and salt over BN254 field elements. The fingerprint and salt stay in the encrypted baseline. The validation path receives the statistical summary and transient capture inputs before the wallet flow submits protocol data on-chain.
06—Prove
06—Prove
Groth16 ZK proof: distance is within the valid range. Not too similar (replay), not too different (imposter).
The circuit proves that both commitments open to the supplied fingerprints. It also proves that their Hamming distance is below the maximum and at or above the replay floor. The circuit does not prove how the client produced either fingerprint.
07—Verify
07—Verify
Proof verified on Solana. Anchor updated. Progressive Trust Score recalculated from verification history.
The private validation service applies phrase, synthesis, capture, and cross-wallet checks to the submitted evidence. Re-verification also requires an on-chain proof. The service processes phrase audio in memory and receives no raw motion or full-resolution touch stream. On success, the Anchor stores the latest timestamp and recent history. Trust Score recalculates from active weekly bins and account age.
// PROTOCOL COMPONENTS
Five layers, one proof.
// FIGURE 01
01
Pulse SDK
Client-side capture and proving
A TypeScript library that captures sensor data, extracts 308 features, creates commitments, and produces ZK proofs. Raw motion and full-resolution touch stay on the device. Phrase audio and derived evidence go to the validation path.
- Browser and React Native support
- Audio, IMU, and touch capture in parallel
- On-device Groth16 proof generation
- Wallet-adapter integration for one-call verification
// FIGURE 02
02
ZK Circuit
Hamming distance verification
A Groth16 circuit that proves two Poseidon commitments open correctly and their Hamming distance falls inside the configured range. It does not prove capture provenance.
- Groth16 over BN254 curve
- Poseidon hash for ZK efficiency
- On-chain Groth16 verification
- Proof generation targets under 5 seconds on mobile
// FIGURE 03
03
On-Chain Programs
Three Solana programs
The entros-verifier program checks ZK proofs. The entros-anchor program stores Trust Score and manages non-transferable Token-2022 Anchors. The entros-registry stores protocol configuration, treasury state, and validator-registration scaffolding.
- Anchor framework with full constraint validation
- Non-transferable token via Token-2022 extension
- Trust Score from active weekly bins and account age
- PDA-derived identity (one per wallet)
// FIGURE 04
04
Executor Node
Off-chain relay and challenge service
A Rust gateway that issues challenges, authenticates integrators, applies quotas, forwards evidence to the private validator, relays protocol writes, and attempts SAS issuance.
- Server-generated signed challenges (anti-bot)
- Best-effort SAS attestation issuance
- Per-integrator API-key rate limiting
- Configurable CORS and per-IP throttles
// FIGURE 05
05
Validation Service
Proprietary defense layer
A private Rust crate that analyzes the 308-dimensional statistical feature summary for synthetic artifacts, submitted-signal statistics, and Sybil patterns. Protocol behavior remains open and auditable. Detection thresholds and model internals stay private because they provide calibration data to an attacker.
- TTS and synthetic data detection
- Cross-wallet fingerprint registry (Sybil detection)
- Cross-signal research telemetry (devnet)
- Uniform public failure classes
// PRIVACY MODEL
Privacy by architecture.
// DATA FLOW
On-device processing
Sensor capture, feature extraction, hashing, and proof generation run on the user's device. Raw motion and full-resolution touch recordings stay in device memory. The spoken phrase leaves the device for validation and transcription.
No raw biometric storage
The validation service processes phrase audio in memory and does not write it to logs or persistent storage. The SDK stores the fingerprint, salt, commitment, and timestamp locally for re-verification. Wallet-connected flows can also store that baseline in a wallet-keyed AES-256-GCM blob on-chain.
Minimal data transmission
The Pulse SDK sends the 308-feature summary, selected F0 and acceleration series, phrase audio, capture timing, client signals, a coarse curve outline, commitment data, and receipt intent to the validation path. The wallet flow submits commitments, proofs, public inputs, and encrypted baseline material on-chain when available.
No identity mapping
The protocol is designed to prove humanness without identifying the person. The protocol does not require a name, email, document, face scan, or social account.
One-way commitment
Poseidon commits to the fingerprint and a large random salt. Its preimage resistance and the hidden salt prevent direct recovery from the commitment under the protocol's threat model.
Data minimization by design
The architecture limits raw-data movement and separates transient validation inputs from persistent protocol state. Each deployer must still assess its own legal and regulatory obligations.
// SECURITY MODEL
How Entros resists bots.
The private service evaluates each submitted capture before settlement. Detection logic stays private. Public campaign results state their tested attack class and denominator.
01
Minimum Distance Constraint
Re-verification proves that committed fingerprints satisfy the circuit's minimum and maximum Hamming-distance bounds. The circuit rejects identical committed fingerprints.
02
Server-Side Feature Validation
The validation service checks the 308-dimensional statistical feature summary before accepting the on-chain proof. It measures acoustic artifacts, submitted-signal statistics, and duplicate fingerprint patterns. Detection logic and thresholds remain private calibration material.
03
Progressive Trust Score
Trust Score uses active weekly verification bins and account age. Repeating a verification inside one weekly bin does not add another active bin.
04
Per-Session Randomness
Each session generates a fresh random phrase and Lissajous curve. The server checks the transcript against the issued phrase before settlement.
05
Multi-Modal Capture
The client captures microphone, pointer or touch, and available motion data. Entros is measuring which combined signals add reliable separation across people, devices, and synthesis methods.
06
Cross-Wallet Fingerprint Registry
The private service keeps a bounded registry of recent server-side fingerprints. It compares new submissions with other wallets under the configured policy. Population-scale performance remains under evaluation.
07
Economic Disincentives
The configured fee and rate limits bound request volume. They do not decide whether a capture passes. The private validation policy makes that decision.
// VERIFICATION FLOW
Behavioral identity state on Solana.
Traditional captcha answers “is this session human?” Entros is designed to measure whether a returning human operates the wallet over time. The protocol provides the verification state. The integrator sets the policy.
// WALLET FLOW
Wallet-connected verification
Connect a Solana wallet. Its Entros Anchor is a non-transferable Token-2022 account. The fingerprint and salt can stay inside a wallet-keyed AES-256-GCM baseline blob. Commitments, Trust Score, and protocol state remain readable on-chain. Each wallet pays transaction fees and account rent for its Anchor.
// RESEARCH VALIDATION
Evidence behind the research program.
External studies motivate multi-modal authentication and synthesis research. Their accuracy numbers do not establish Entros performance on consumer devices.
Mobile fusion can improve authentication
BioMoTouch, arXiv 2025
BioMoTouch reports 99.71% accuracy and a 0.27% equal error rate for touch and motion fusion in its evaluation. This motivates Entros research, but does not establish Entros liveness or resistance to generated input.
Arm movement can perturb the voice
Pouw et al., Royal Society Proceedings B, 2025
Pouw et al. found that prompted upper-limb movements and related postural muscle activity changed the amplitude envelope of a sustained /a/ vowel under controlled laboratory measurement. Whether consumer devices can turn that effect into a reliable liveness signal remains an open Entros research question.
Physics-informed features merit evaluation
VoiceRadar, NDSS 2025
VoiceRadar reports a 0.45% equal error rate for its micro-frequency feature on its benchmark. This motivates Entros evaluation but does not establish an Entros error rate.
Modern TTS produces artifacts in both directions
Warren et al., "Pitch Imperfect", 2025
Neural vocoders can produce acoustic perturbation values higher than human baselines, not just lower. This overturns the assumption that synthetic speech is "too perfect." Detection must account for artifacts in both directions.
Touch and motion can share a physical response
Device physics, BioMoTouch 2025
A physical touch can produce a time-aligned response in device motion sensors. Whether consumer devices and attested clients can measure that response reliably enough to distinguish physical interaction from generated input remains open research.
// CORE ADVANTAGE
Entros captures voice, movement, and touch in one session. The current fingerprint projection mixes those features but does not prove a causal relationship between them. Temporal analysis stays observable while challenge-bound designs and human completion are tested. Trust Score records the wallet's verification history.