IIC 3.0 (International Innovation Challenge, Manipal University Jaipur) — Theme: Cybersecurity & Digital Sovereignty
India's largest quantified fraud is the digital-arrest scam: fake CBI/police calls that coerce victims — mostly elders — into transferring their life savings. ₹4,057.7 crore lost across 297,727 complaints (2022–May 2026). Losses grew 20× by 2024. Scammers now pair coercion scripts with AI-cloned voices ("Dad, I'm in jail, send money") — a clone takes 30–60 seconds of harvested audio, costs pennies, and the call arrives with a spoofed caller ID.
Every bank warns customers. No consumer app defends the phone itself — in Hindi.
Kavach is a call-security platform that fuses six detection departments into one intervention loop:
| Department | Question |
|---|---|
| Spoof | Is this voice real? (AASIST-hindi, 3-crop vote) |
| Coercion | Is this a scam script? (Hindi ASR + 8 vector banks) |
| Threat | Is this caller dangerous? |
| Factcheck | Is this claim true? |
| Payment Tripwire | Is a payment about to happen? (before the PIN) |
| Evidence + Report | Tamper-proof packet → 1930 / Chakshu |
The FUSION core turns all signals into one verdict → one intervention: recognize → interrupt → verify → package → report.
The product loop: the moment a call shows both spoof AND coercion signals, Kavach pauses the payment moment — warns the victim in Hindi, alerts a trusted family member, and generates a tamper-proof evidence packet ready for 1930.
call audio → mel-spectrogram → anti-spoof model (AASIST-hindi, 3-crop majority vote) → spoof score
→ faster-whisper Hindi ASR → coercion state (8 vectors) → FUSION verdict → PAUSE → evidence → 1930
- Model: AASIST (Attention-based Spectrogram Transformer) fine-tuned on Hindi deepfake data — 0.9919 accuracy, 0/100 false positives on held-out Hindi spoof data (re-verified 2026-08-11)
- Latency: ~71 ms mean full-pipeline (3-crop vote, max ~100 ms) on a GTX 1650 (4 GB) laptop GPU — on-device scoring
- Honest platform note: consumer Android apps cannot record live cellular calls (VOICE_CALL is privileged). Kavach owns the moments Android permits: pre-ring screening, own-mic analysis burst, the payment tripwire (Play-sanctioned Accessibility for fraud prevention), the evidence capsule, and family escalation. Higher-assurance tiers (CPaaS second number, bank/FRI partnership) are roadmap.
v1_loop.py — THE executable: full product loop (recognize→fuse→interrupt→package→report)
run: python v1_loop.py <audio> [--payee-new] [--amount N] [--decision approve|challenge|kill]
python v1_loop.py --demo # 5-scenario battery
src/engine.py — KavachEngine: AASIST-hindi model load, 3-crop vote, analyze()
src/coercion.py — Hindi coercion detection (8 vectors, fuzzy, rule boosts)
src/fusion.py — FUSION core: 6 departments → one verdict ladder (PASS/CAUTION/PAUSE/KILL)
src/evidence.py — sha256 chain-of-custody packet (KV-) + 1930-ready PDF
demo_ui.py — Gradio UI: upload/record → verdict + evidence packet
demo/b3-intervention-mock.html — PAUSE screen + family-challenge UI flow mock (static preview)
benchmarks/ — the test battery (all re-runnable): audio 31/31 · stress 226/226
· real-cases 29/29 (24 documented incidents) · curated 10/10 · D7 90%
scripts/ — capture + verify scripts
assets/ — screenshots, evidence packets, sample audio, deck
models/ — aasist-hindi.pt + AASIST weights (tracked)
b1_verify.py — smoke test (weights load, CUDA, latency)
b1_ab_test.py — A/B evaluation script
b2_finetune.py — Hindi fine-tuning script
| Input | Verdict | Spoof score | Latency |
|---|---|---|---|
| Hindi voice-clone attack (edge-tts) | SPOOF → PAUSE | 1.000 | ~70–100 ms |
| Real Hindi call audio (FLEURS test set) | BONAFIDE → PASS | 0.000 | ~70–100 ms |
Full benchmark battery (2026-08-11, all re-runnable): audio 31/31 · stress 226/226 · real-case registry 29/29 · curated 10/10 · evidence mutation 90% · unit 31 OK. The real-case registry replays 24 documented Indian scam incidents (BBC/NDTV/IE/TOI/FPJ) — every documented script is flagged. Evidence packets: every detection emits a sha256-chained JSON audit trail + 1930-ready PDF (53 KV- packets, tamper-verified).
# 1. install deps (Python 3.10+; NVIDIA GPU recommended)
uv venv && uv pip install -r requirements.txt
# 2. run the full product loop on any call audio (real or synthetic)
python v1_loop.py <call.wav|mp3> [--payee-new] [--amount 150000]
# 3. run the 5-scenario demo battery (3 attacks + 2 real calls)
python v1_loop.py --demo
# 4. launch the Gradio UI (upload/record → verdict + evidence packet)
python demo_ui.py # → http://127.0.0.1:7860
# 5. re-run the test battery
python benchmarks/run_audio.py && python benchmarks/run_stress.py && \
python benchmarks/run_real_cases.py && python benchmarks/run_mutation.pyModels ship in models/ (aasist-hindi.pt + AASIST weights) — no download needed. The first v1_loop.py call downloads the faster-whisper small (hi) ASR model (~460 MB, cached).
- ✅ B1 spoof engine — built, verified (0.9919 acc, 0/100 FP, re-verified 2026-08-11)
- ✅ B2 coercion layer — built (Hindi ASR + 8 vectors; audio suite 31/31, stress 226/226, real-case registry 29/29)
- ✅ B4 evidence packet — built (sha256 chain, D7 mutation 90% caught, 53 KV- packets)
- 🟡 B3 intervention UI — UI-flow mock (static preview in
demo/), full Android build is R2
- RBI/2024-25/105 — voice/SMS fraud safeguards
- CERT-In CIAD-2024-0060 — Deepfakes: threats & countermeasures
- DoT FRI: 1000+ banks · ₹660 crore prevented (Dec 2025)
- SC order (Aug 4 2026) — digital-arrest complaints 1,23,672 → 16,377 (I4C data)
- ASVspoof 2021 LA — anti-spoofing benchmark + baselines
- AASIST — clovaai/aasist (github.com/clovaai)
GenAI used for ideation/support only. Model training, evaluation and this submission's technical claims are our own work.