AuthMark: Cryptographically-Verifiable Media Authentication

September 2024 – August 2025

I designed, and implemented, a novel and proactive system to directly address the "Erosion of Trust" in digital media, by providing a cryptographic link between the human author and their work.

• Engineered a hybrid, Python-based system, that embeds a SHA-256 hashed, and a Ed25519 (EdDSA) digitally-signed payload into an image's Discrete Cosine Transform (DCT) mid-frequency coefficients.
• Achieved superior quantitative watermark imperceptibility, validated by an average Peak Signal-to-Noise Ratio (PSNR) of 47.06dB, and a Structural Similarity Index (SSIM) of 0.994 – far greater than the accepted industrial threshold of object imperceptibility.
• Built a robust framework resilient to common image manipulations, achieving 100% watermark recovery against JPEG compression down to a Quality Factor (QF) of 55%.
• Integrated a hierarchical AI detection workflow using "DMImageDetection", a model openly available on Hugging Face, to proactively mitigate the issue of undesired synthetic imagery watermarking.
• Deployed as a lightweight web application, with a comprehensive RESTful API, using Flask, and Flask-RESTX, prioritising accessibility, ease of integration, and hypothetical vast scalability to other web platforms with varying purpose and available computational resources.
• Validated the system's real-world efficacy through a ten-person blind study, whereby its 95.92% accuracy in identifying authentic media surpassed average human perception.