Deep-Live-Cam: AI Content Authentication System

closed source
Oct 2025 - Dec 2025
PythonOpenCVPyTorchDockerRSA-2048

AI video/img LSB watermarking security layer using RSA-2048 and stenography

Built a dual-layer security system for a deepfake application using invisible LSB watermarking, RSA-2048 digital signatures, and a stenography module to embed cryptographic metadata into image pixels surviving JPEG compression. Developed PKI infrastructure with automated key generation, AES-256 encrypted private keys, and signature verification tools; containerized the ML pipeline with Docker for cross-platform use.

Content authentication for deepfake detection via watermarking and PKI. • Watermarking: Invisible LSB embedding and stenography module so metadata survives JPEG compression. • PKI: RSA-2048 digital signatures; automated key generation; AES-256 encrypted private keys; verification tools using Python's cryptography library. • ML: ONNX Runtime, InsightFace, OpenCV, PyTorch; pipeline containerized with Docker for cross-platform use.

Case Study

Problem

Deep-fake video tools generate convincing synthetic content with no built-in provenance. Add an authentication layer that lets anyone verify whether a frame was produced by a specific model instance.

Architecture

  • LSB steganography module embedding cryptographic metadata invisibly into pixel channels
  • Compression-resistant payload design (survives JPEG re-encode up to ~85% quality)
  • RSA-2048 digital signature generation and verification using Python cryptography library
  • AES-256-CBC encrypted private key storage for PKI key material
  • CLI tools for key generation, signing, and verification
  • ONNX Runtime + InsightFace + OpenCV + PyTorch ML pipeline; Docker-containerised for portability

Challenges

  • Ensuring LSB payload survived JPEG compression (chose bit planes 0–1 and channel weighting)
  • Keeping RSA signature generation below 40 ms per frame at 30 fps
  • Designing a key-derivation scheme that is reproducible yet resistant to brute force
  • Cross-platform Docker image for both CPU (x86) and GPU (CUDA) inference

Tradeoffs

  • LSB steganography is invisible but fragile against aggressive re-encoding; balanced by targeting bit planes least affected by DCT
  • RSA-2048 chosen over ECDSA for wider tooling compatibility despite larger key size
  • Closed-source to protect proprietary watermarking algorithm design

Outcome

Dual-layer security system passed all synthetic test cases; watermark survived JPEG compression at 85% quality and signatures verified end-to-end.

What I Learned

  • How LSB steganography interacts with DCT-based compression codecs
  • Python cryptography library internals: RSA, AES key derivation, padding schemes
  • ONNX Runtime inference optimisation for real-time video pipelines
  • PKI design patterns: certificate chains, key revocation, and secure storage