MLWM v1 adds a neural short-payload image watermark engine next to the legacy DCT engine.
- Payload is fixed to 256 bits after framing, RS encoding, and deterministic bit whitening.
- Text payload is limited to 16 UTF-8 bytes.
- Legacy frequency-domain synchronization remains the geometric anchor.
- Neural embed/extract only handles payload recovery.
- Runtime uses ONNX for desktop inference and PyTorch for training/export.
- Encode short text into a fixed frame with CRC32 and RS parity, then whiten the 256 training bits.
- Run the encoder network to predict a residual map.
- Apply the residual to the image and inject the classical sync template.
- At extraction time, detect the sync template, rectify the image, and score several candidate views.
- Aggregate decoder logits, unwhiten the bits, RS-decode, and CRC-check the payload.
- If neural decode fails in
auto, fall back to the legacy engine.
- Stage A: decoder warmup with deterministic fixed residuals.
- Stage B: joint encoder/decoder training with medium attacks.
- Stage C: hard-negative fine-tune on the strongest attack chains.
- Stage D: confidence calibration and export freeze.