Visual

What it processes

Classifies image and video content frame by frame. Looks for unsafe imagery, encoded payloads carried in imagery, and visual content engineered to read benign to single-shot models.

Adversarial patterns it is tuned to catch

• Pixel-encoded prompts and steganographic payloads
• Adversarial perturbations tuned to flip a classifier verdict
• Frame-level evasion in otherwise benign video
• Visual abuse content disguised through filters and overlays

Contribution to the ensemble verdict

Per-frame and per-clip scores feed a temporal-aware aggregator before contributing to the ensemble verdict.

Per-detector outputs are not a final verdict. The ensemble layer reads all nine and decides; the per-detector contribution is preserved in the evidence trace so the verdict remains auditable.

Related research

• Adversarial content classification: lessons from operating at web scale
• Cross-modal evasion and how to defeat it

All nine detectors See the pipeline