SHANSHIELD

SHANSHIELD is an agentic, multi-modal deepfake detection platform that analyzes images, video, audio using 6 specialized AI agents working in concert with a option to run 100% in the browser.

Comment

SHANSHIELD — Agentic AI Deepfake Defense System( LIVE ON https://shanz.co.in ) 💡 Inspiration

Deepfakes are evolving faster than defenses. Voice cloning now takes 3 seconds. Existing tools are single-modal, cloud-dependent, and black-box — unusable in field conditions (courtrooms, conflict zones, newsrooms). We asked: what if detection was multi-agent, offline-first, explainable, and quantum-secure?

🛠️ How We Built It

Architecture: A 6-agent pipeline where each agent specializes in one forensic domain, that uses GEMINI 3 API Cloud ML model and also option to run entirely in-browser:

Agent Domain Core Algorithm Visual Pixel forensics Sobel edge convolution, noise variance ($\sigma^2$) Audio Spectral analysis FFT, pitch autocorrelation Temporal Frame consistency Inter-frame differencing AI Signature Tool fingerprinting Pattern matching (Midjourney, Sora, etc.) Metadata File structure EXIF/header analysis Quantum Entropy Randomness profiling Von Neumann entropy: $S = -\text{Tr}(\rho \log_2 \rho)$ Evidence Fusion: The Arbiter combines all agent outputs using Dempster-Shafer theory:

$$m_{1,2}(A) = \frac{\sum_{B \cap C = A} m_1(B) \cdot m_2(C)}{1 - \sum_{B \cap C = \emptyset} m_1(B) \cdot m_2(C)}$$

This avoids single-point-of-failure and provides mathematically grounded confidence scores.

Cloud Cross-Verification: Google Gemini 3 (gemini-3-flash-preview) provides a second opinion via Lovable Cloud backend functions, with hybrid scoring:

$$\text{Score}{\text{final}} = \alpha \cdot S{\text{offline}} + (1 - \alpha) \cdot S_{\text{cloud}}$$

Post-Quantum Integrity: Every verdict is signed with CRYSTALS-Dilithium3 (NIST FIPS 204), ensuring tamper-proof provenance even against quantum adversaries.

Stack: React + TypeScript + Vite + Tailwind CSS, with browser-native Web Crypto API (SHA-256), jsPDF forensic reports, and C2PA content authenticity verification.

📚 What We Learned

Dempster-Shafer fusion is far more robust than naive averaging — it handles conflicting evidence gracefully Browser-native forensics (Canvas API, Web Audio FFT, Web Crypto) are surprisingly powerful for real-time analysis Quantum entropy ($S_{\text{vN}}$, Rényi entropy $H_\alpha$) is a genuine differentiator — AI-generated content exhibits measurably lower entropy than natural media Designing for EU AI Act and GDPR Article 22 compliance from day one forces better architecture decisions (explainability, no data storage) 🚧 Challenges

Every-frame video analysis in-browser: sequential seeked event processing at 30fps without crashing the tab required careful memory management Dempster-Shafer with 6 agents creates combinatorial complexity — we optimized by pre-filtering low-mass hypotheses PQC in JavaScript: CRYSTALS-Dilithium3 has no mature browser library — we implemented a simulation aligned to NIST specifications Balancing offline vs cloud: ensuring the system is fully functional offline while gracefully enhancing with Gemini 3 when connected Explainability: making every verdict traceable to specific pixel regions, audio segments, and per-agent reasoning — not just a confidence number 🏆 Awards

🥇 Winner — IIT Delhi E-Raksha Hackathon 🥇 Winner — ASME IMECE 2025 Innovation Pitchathon

Built With

  • browser-native-forensics-?-crypto:-dilithium3
  • dempster-shafer-?-standards:-c2pa
  • eu-ai-act
  • frontend:-react
  • lovable
  • sha-256-?-math:-fft
  • sobel
  • tailwind
  • typescript
  • vite-?-ai:-gemini-3
  • von-neumann-entropy
Share this project:

Updates