Why this matters now
Watermarking schemes like SynthID are becoming a standard for provenance in AI-generated imagery, but they also spark a security and robustness research cycle. This repository shows a counterpoint: using only signal processing and cross-image statistics (no access to Google's encoder), it reveals SynthID's resolution-dependent carrier structure and an invariant phase template—insights that let researchers detect the watermark (~90% detector accuracy) and perform targeted spectral subtraction (V3) with ~43 dB PSNR and a ~91% phase-coherence drop on evaluated images.
What Sets It Apart
- Multi-resolution, codebook-driven approach: instead of blind denoising, the project builds per-resolution fingerprint profiles (from pure black/white references and watermarked sets) so removals operate on exact FFT bins. So what? That enables surgical subtraction with far less visual degradation than brute-force transforms.
- Phase-aware subtraction and confidence weighting: carriers are validated by cross-image phase coherence and black/white agreement; subtraction is scaled by confidence and capped per-bin. So what? It reduces false positives and preserves image content while effectively suppressing watermark energy.
- Empirical metrics and reproducible profiles: reported detector ~90% accuracy; V3 results show ~75% carrier energy drop and ~91% phase coherence drop across tested resolutions, with PSNR > 43 dB on aggressive settings. So what? The results quantify a realistic arms-race scenario between watermark design and signal-processing removal.
Who It's For and Trade-offs
Great fit if you are a researcher, security analyst, or watermarking designer who needs a concrete adversarial benchmark against SynthID-style spread-spectrum watermarks, or if you study robustness across image resolutions. The project is explicitly research-oriented: its methods assume access to many reference images (black/white Gemini outputs) and a curated codebook per resolution, so performance degrades without those references.
Look elsewhere if your goal is practical, production-grade content moderation or provenance enforcement—this repo demonstrates how a determined analyst can reduce watermark detectability and therefore highlights limitations of the watermarking scheme, but it is not a turnkey moderation system. The README and code include an ethical disclaimer; do not use these tools to conceal AI authorship.
Where It Fits
Compared with brute-force quality degradations (JPEG, heavy noise), this work shows that targeted frequency-bin subtraction guided by cross-image phase statistics achieves much higher visual fidelity for a given level of watermark suppression. Treat this repo as a defensive/academic benchmark that surfaces specific failure modes of per-image watermarks tied to fixed carrier patterns and resolution dependencies.
