Watermark Robustness

Watermark robustness describes how well an invisible AI-content watermark survives real-world transformations — resizing, recompression, format conversion, re-uploading to social platforms — versus deliberate, targeted removal. A robust watermark is built to survive the first category; no publicly known watermarking scheme, including Certivu's, is claimed to survive the second. Understanding that distinction is central to using watermarks honestly as a resilience signal rather than as tamper-proof evidence.

What watermarks are designed to survive

Certivu's watermark embeds an opaque lookup identifier into mid-frequency regions of an image or audio signal using frequency-domain (DCT) techniques, chosen specifically because ordinary processing — JPEG recompression at reasonable quality, resizing, and the transformations social platforms apply automatically on upload — tends to preserve those frequencies. That's what makes a watermark useful as a fallback: when a token or embedded manifest gets stripped by a platform's re-encoding pipeline, the watermark can often still be recovered and matched back to a signed provenance record.

What watermarks are not claimed to survive

Watermarking is not claimed to be unremovable. Someone specifically motivated to strip a watermark — through heavy cropping, adversarial noise designed to target the watermark, aggressive re-compression, or manual editing focused on defeating detection — can degrade or remove it, a limitation every publicly known image or audio watermarking scheme shares. This is why Certivu treats a watermark as a resilience signal, not proof: the cryptographic signature on a signed record is the actual source of truth, and a missing or degraded watermark doesn't invalidate a valid signature reachable another way, nor does it mean the content was never signed.

Why this honesty matters

Overstating watermark durability creates a false sense of security and, in a compliance context, could mislead a platform or regulator about what a marking actually guarantees. Certivu's verification response reflects this directly: watermark-based matches contribute to a lower confidence tier than a verified cryptographic signature, and the system never claims a watermark's absence proves human origin — it may simply mean the mark was degraded, removed, or never present in the first place.