Confidential Error-Correcting Codes

What it claims

Confidential Error-correcting Codes (CECs) are a cryptographic primitive that decodes from perturbed codewords while keeping the message hidden. The abstract states that the concrete instantiation, Conc-CEC, has provable guarantees for all three properties: confidentiality, robustness and soundness. It also claims "significantly stronger concrete security guarantees" than PRCs (CRYPTO'24) "under practical parameter settings." The application is CECMark, an image watermark for AI-generated content. It is reported to beat Gaussian Shading (CVPR'24), PRCMark (ICLR'25) and T2SMark (NeurIPS'25) on robustness across all 56 evaluated perturbations, with a false positive rate below 10⁻⁵ and "without introducing side effects."

How it works

The construction follows a Randomness-Security Separation (RSS) structure, which splits randomness from secrecy:

1. Encode a nonce, which supplies the randomness.
2. Mask the encoded payload with a pseudorandom function (PRF), which supplies the secrecy.

Conc-CEC builds this from concatenated Reed-Muller and Reed-Solomon codes. The abstract does not say:

• which code is inner and which is outer;
• the code parameters;
• how the PRF key is managed;
• how CECMark embeds the codeword into images, for example in the latent or the noise of a diffusion model.

Limits

• Only the abstract is on the page. Everything above comes from it. Several things the paper itself may contain are not available here:
• the security bounds;
• the parameter choices behind the "stronger concrete security" claim against PRCs;
• the list of 56 perturbations;
• per-perturbation robustness numbers;
• which generative models were tested, described only as "real-world generative image models";
• what "side effects" means and how its absence was measured.
• Comparison not quantified. There is no number for the margin over the baselines, either on robustness or on concrete security.
• No code mentioned. The abstract does not mention a code or artifact release. Adopting CECMark currently means reimplementing it from the paper.
• Status. This is a preprint, received 2026-10-10 and approved on ePrint 2026-10-14. ePrint approval is not peer review.
• Licence. The paper is CC BY. That licence covers the text and says nothing about any implementation.
• Authors and scope. The authors are from Tsinghua and Shandong University, including Xiaoyun Wang. ePrint files the paper under Applications, with keywords watermarking, AIGC and code-based cryptography. The work is aimed at image watermarking; nothing in the abstract addresses text or other modalities.