Model Card: Starlight Unified Model 2025

Model Overview

• Task: Detection / Extraction
• Architecture: Unified CNN-based Encoder-Decoder with Residual Blocks
• Input: 256x256 RGB/RGBA or metadata
• Output:
  • Detector: sigmoid probability
  • Extractor: variable-length byte sequence

Training

• Dataset: Combined submissions (grok, gemini, claude, chatgpt, sample)
• Epochs: 50
• Batch Size: 16
• Optimizer: Adam
• Loss: BCE + MSE (detector), CrossEntropy (extractor)

Performance

Metric	Value
Accuracy	96.3%
AUC-ROC	0.996
F1 Score	0.982
Extraction BER	0.003

Steganography Coverage

• lsb, alpha, dct, exif, eoi, palette

Inference Speed

• CPU: 12 ms/image
• GPU: 2.1 ms/image

License

• Model: Apache 2.0
• Code: MIT