Smart India Hackathon 2026 · Problem Statement 26104

Six slides,
in full

The official six-slide format — title, proposed solution, technical approach, feasibility and viability, impact and benefits, research and references — each carrying the detail a slide cannot hold.

01 Title page

VoxShield — it refuses to guess

An offline voice-cloning detector that treats “I cannot tell” as a real answer — and makes that answer stop a fraudulent transfer just as firmly as a detection would.

Problem Statement ID26104
ThemeBlockchain & Cybersecurity
PS CategorySoftware
Problem Statement TitleAI-Powered Real-Time Detection and Prevention of Voice Cloning Impersonation Attacks
OrganisationAICTE — Cyber Security Cell
Team NameInnovexa
Team ID[PLACEHOLDER]

Copy these fields from the portal character for character. A theme or category that does not match the listing is one of the cheapest ways to lose marks, and no placeholder text should survive into the submitted file.

Team Innovexa · SIH 2026 · Software Project record
02 Proposed solution

The gap is not detection. It is what happens when detection is unsure.

How the attack runs

01Seconds of real voiceA voicemail, a reel, a recorded meeting.
02Cloned by generative AISpeech in that voice, on demand.
03A trusted person callsAn executive, an official, a family member.
04Urgency is manufacturedDo it now, do not check, do not tell anyone.
05Transfer, OTP, accessAuthorised by a voice that was never there.

Published work measured a well-known detector at 3.15% error in the lab and 37.8% on real-world audio. A detector wrong roughly one time in three that still prints “SAFE” is dangerous inside a bank.

Which is why our claim is the decision layer, not a better detector.

Why the existing defences do not close it

ApproachVoice analysisWorks offlineHandles uncertaintyDrives an actionAuditable
Caller IDNoYesNoNoNo
Manual callbackNoYesHuman judgementDepends on the personNo
Recognising the voiceBy earYesNo — familiarity is the attackNoNo
Binary deepfake detectorYesVariesReturns “safe” when unsureScore onlyVaries
Cloud detection APIYesNo — audio leaves the premisesVariesScore onlyProvider-side
VoxShieldYesYes — zero outbound calls at runtimeExplicit refusal with a reason codePrescribed action per risk profileHash-chained log

What makes it different — three things

The gate runs first

Built

If there is too little clear speech or the line is too poor, the model score is never consulted. The answer is LOW_CONFIDENCE, never SAFE.

Uncertainty is a number

Built

“Risk 55–90, cannot narrow further: only 1.4 s of clear speech” — not a vague hedge that helps nobody decide.

Uncertainty has a consequence

Built

At high-value transfer and privileged access, a refusal and a detection produce the same action: hold and call back.

The sentence for the slide: no failure anywhere in this system — short audio, a bad line, a model crash, a GPU overload — can produce a SAFE verdict.

Detection + decision, not detection alone What we claim, and do not
03 Technical approach

Named components, one pipeline, four layers

The ecosystem

Built Gate — before any model score is read Planned

Named technology, with licences

LayerNamed technologyLicence
FrontendPlain HTML, CSS, vanilla JS console · Kotlin + Jetpack Compose on Android
ServingFastAPI + uvicorn, single worker, bound to localhostMIT
StorageSQLite (WAL), no ORM · on-device SQLite for the appPublic domain
Voice activitySilero VAD (ONNX)MIT
Speaker identityECAPA-TDNN — speechbrain/spkrec-ecapa-voxcelebApache-2.0
Synthesis artefactsAASIST (console) · RawNet2 via ONNX Runtime Mobile (app)MIT
Language checkSpeechBrain language-ID model — routing only, never a coverage claimApache-2.0
Phase + prosodylibrosa + NumPy — built, measured, deliberately excluded from fusionISC / BSD
IntegritySHA-256 hash-chained audit log · SHA-256 model verification at load
Telephony interfaceSIPREC (RFC 7866) · carrier media-stream WebSocket formatIETF standard

Detection pipeline

01Ingest

WebSocket or file.

0216 kHz mono

Frozen wire format.

034 s window

1 s hop — a verdict a second.

04Silero VAD

Speech or silence.

05Precondition gate

Before any score.

06Language check

Routing only.

07Validation gate

Answer only where checked.

08Signals

Synthesis + speaker.

09Fusion

Weighted, bounded range.

10Risk 0–100

Or a range plus a reason.

11Verdict

Five outcomes.

12Audit, discard

Window leaves memory.

The precondition gate

Speech≥ 2.5 s voiced
Line qualitySNR ≥ 10 dB
Clipping≤ 2%
EnrolmentA voiceprint to compare with

Fail any one and the score is never read: LOW_CONFIDENCE with one of six reason codes.

Verdict spectrum

SAFE0–39
SUSPICIOUS40–69
HIGH_RISK70–100
LOW_CONFIDENCErange + reasonis_fake: null · not SAFE

Risk profile drives the action

Risk profileSUSPICIOUSHIGH_RISKLOW_CONFIDENCE
Routine enquiryLogFlag to agentLog
Account changeFlag to agentHold + callbackPrompt secondary question
High-value transferHold + callbackBlock + escalateHold + callback
Privileged accessBlock + escalateBlock + escalateBlock + escalate

The action is prescribed and displayed today; automated execution — incidents, step-up verification, alert delivery — is specified and not yet built, and is labelled that way wherever it is demonstrated.

One process · one CUDA context · ~582 MB resident of a 6 GB card Full technical breakdown
04 Feasibility & viability

Name the weak points before a juror does

It is what makes the rest of the deck believable — and it is exactly what the product itself argues.

308 msEnd-to-end checkRTX 4050 laptop GPU, 4 s window.
64Calls at once0 dropped. Where testing stopped, not where it breaks.
0Bytes of audio storedNot to disk, not to a log, not to a temp file.
₹0Build cost21 dependencies, all permissive, licence-audited in the build.

Feasibility on five axes

TechnicalPretrained models only — no training, no fine-tuning, no dataset licensing. ~582 MB resident on a ₹60k-class laptop GPU.
OperationalRisk profiles, prescribed actions and human review fit an existing fraud-operations workflow.
Economic₹0 total cost. No cloud contract, no per-minute fee, no API key.
ScalabilityOne process, one CUDA context, bounded queue. 64 calls measured; capacity reported as a floor, never a ceiling.
PrivacyMemory-only audio, encrypted voiceprints, deployable inside an institution's own perimeter.
DeploymentFully offline after install — no internet, no account, nothing to provision.

Challenge → engineering response

ChallengeWhat it changed in the architecture
Detectors degrade on real-world audioThe precondition gate exists because of this. The system degrades to a refusal, never to a confident wrong answer.
A detector may be unreliable on audio it was never tested onA validation gate withholds the signal and returns NO_VALIDATED_SIGNAL rather than reporting a score we do not believe.
Signals can disagreeA disagreement beyond the threshold is reported as SIGNAL_DISAGREEMENT, not averaged into a middle number.
Phones block call capture at OS levelMeasured on a real device rather than assumed; the enterprise path attaches to infrastructure the institution already owns (SIPREC or a carrier media stream).
No budget for a telephony accountImplemented against the documented protocol and tested against a local mock that speaks the same format.
GPU memory is finiteOne process and one CUDA context with a bounded queue — extra processes would multiply memory and serialise anyway.
A model file could be silently swappedEvery weight file is pinned by SHA-256 and verified at load; a mismatch aborts startup.
Fusion weights are not calibratedStated openly as a rule of thumb. Calibration with proper cost reporting is named as the next version, not implied as done.

Compliance position. The DPDP Rules were notified on 14 November 2025. A voice embedding that identifies a person is personal data, not anonymised features — so voiceprints are encrypted at rest, never returned by any API, and destroyed cryptographically on erasure, with every creation and deletion audited.

351 tests · 91.97% coverage · 15/15 model files verified Measurements & limits
05 Impact & benefits

Uncertainty becomes a decision instead of a gap

Who benefitsHow
Bank fraud operationsA live verdict with the reason and a prescribed action — and a rule for the calls where the system cannot tell.
CustomersFewer successful impersonation frauds, and no voice data leaving the institution.
Security architectsNo new capture surface: the enterprise path reads media the session border controller already forks.
Regulators and auditorsA tamper-evident decision trail where an altered entry is named by the verify step.
Smaller institutions₹0 software on commodity hardware — no cloud contract, no per-minute fee.
Individuals and familiesAn on-device warning during the call, when a pause still costs nothing.

Where the impact lands

Social

Protection from impersonation scams for the people targeted most, and a reason to verify that does not depend on suspicion.

Economic

Reduced exposure to fraudulent transfers at the moment of the request, with no recurring licence or cloud cost.

Educational

The method, the failures and the limits are published with the work — it can be checked, repeated and taught.

Technological

A reusable pattern for high-stakes AI: gate first, refuse explicitly, and give the refusal a consequence.

Operational

Four risk profiles mean a balance enquiry and a large transfer are not held to the same strictness.

Environmental

Inference runs on hardware already owned, with no cloud round trip per call. We claim nothing beyond that, because nothing beyond that has been measured.

Where it is used

BankingA transfer requested by phone → hold and call back rather than approve on a voice
EnterpriseExecutive impersonation to finance or the help desk → escalate instead of comply
TelecomVoice security at the carrier media stream, without recording anything new
GovernmentIdentity impersonation of officials → a recorded, reviewable decision
ConsumerA familiar voice asking for money urgently → an on-screen warning during the call
After the factA structured record for a complaint on the national cybercrime portal
No audio leaves the device or the premises Security & privacy
06 Research & references

Real sources, with links

Standards and official sources

SourceUsed forLink
RFC 7866 — SIPREC Session Recording ProtocolThe enterprise telephony interface we implement against.rfc-editor.org
RFC 6962 — Certificate TransparencyThe Merkle design a later version of the audit log would move to.rfc-editor.org
Problem Statement 26104, SIH 2026The statement this project answers.sih.gov.in
AICTE — Cyber Security CellThe organisation behind the problem statement.aicte-india.org
Digital Personal Data Protection Act, 2023 and its Rules (notified 14 Nov 2025)Why a voiceprint is treated as personal data, encrypted and erasable.meity.gov.in
National Cyber Crime Reporting PortalWhere a user files the complaint our record export is prepared for.cybercrime.gov.in

Research papers

PaperKey point usedLink
Does Audio Deepfake Detection Generalize? (In-the-Wild)The lab-versus-real-world gap that the whole decision layer answers.arXiv:2203.16263
AASISTSpectro-temporal graph attention for anti-spoofing — the synthesis signal in the console.arXiv:2110.01200
ECAPA-TDNNSpeaker verification architecture behind the identity signal.arXiv:2005.07143
RawNet2End-to-end anti-spoofing from raw waveform — the on-device signal in the app.Paste the exact paper link used — [PLACEHOLDER]

Models and datasets

ResourceRoleLink
Silero VADVoice activity detection, ONNX, MIT.github.com/snakers4/silero-vad
speechbrain/spkrec-ecapa-voxcelebThe speaker-identity model card and weights.huggingface.co
speechbrain/lang-id-voxlingua107-ecapaLanguage identification, used to route rather than to claim coverage.huggingface.co
ASVspoofBenchmark and protocols for spoofed-speech detection.asvspoof.org
FLEURSGenuine multilingual speech used in our own test runs.huggingface.co
MLAAD / other spoof corporaSynthetic speech used in our own test runs.Paste the exact dataset link — [PLACEHOLDER]

Roadmap — and what it is not

Now

Built

Detection pipeline, precondition gate, verdicts and profile actions, audit chain, sign-in and roles, grounded assistant.

Next

Specified

Incidents and step-up verification, alert delivery, phone-to-console hand-off, calibrated fusion weights, SIPREC recording server.

Future

Planned

Wider validated coverage on evidence, Merkle inclusion proofs, SDK, cross-platform and carrier/OEM integration.

Know more about VoxShield. This site is the long version of the deck: the full walkthrough, the technical breakdown, the measurements and limits, the current build stage and the team.

No source is listed here unless it supports a claim we actually make Documentation hub

Know more about VoxShield

Voice → verify → risk → act → audit

VoxShield does not claim to always know. It claims to know when it does not know — and to act accordingly.