Imagine you want to prove that an image, a piece of text, or a software program belongs to you, or was generated by your AI model. The traditional way to do this is through symmetric digital watermarking: you use a secret key to embed a hidden pattern into the asset, and you use that exact same secret key to detect it later.
This creates a fundamental security paradox. If a third party—such as a social media platform, a regulatory body, or an end-user—wants to verify the watermark, you must give them the key. But because the system is symmetric, anyone who holds the verification key also gains the power to completely remove the watermark or forge it onto unauthorized content.
To solve this dilemma, researchers have developed publicly verifiable (or asymmetric) watermarking. The core intuition is to decouple the embedding process from the verification process. By using two different keys—a private key for embedding and a public key for detection—anyone can audit and verify the provenance of an asset without ever gaining the ability to alter or fake the mark.
The Vulnerability of Shared Secrets
To understand why asymmetric watermarking is necessary, we must first look at how easily symmetric systems break down under adversarial scrutiny.
When the same key governs both embedding and detection, the security of the entire system hinges on absolute confidentiality. As Duhamel et al. pointed out in Introduction aux méthodes de tatouage asymétriques dans le cadre de la protection de copie, symmetric methods (especially correlation-based or spread-spectrum techniques) are highly vulnerable to collusion attacks. If an attacker collects multiple watermarked instances generated with the same key, they can simply average them together to isolate and reverse-engineer the underlying secret structure.
Furthermore, modern attacks do not even need to guess the exact key to neutralize a watermark. As demonstrated by Lukas et al. in Leveraging Optimization for Adaptive Attacks on Image Watermarks, many watermarking schemes rely on a consistent embedding channel that does not vary significantly with the key. This allows attackers to bypass security entirely by optimizing against a single "surrogate key," rendering the secrecy of the original key useless.
This fragility is a shared trait among many key-dependent defenses. Whether protecting image classifiers, diffusion models, or linear algebra backends, if an adversary compromises or approximates the secret key, they can reconstruct the gradient landscape to remove the protection or execute white-box attacks (Taran et al.; Yang et al.; Möller et al.).
Symmetric Watermarking (Fragile):
[Private Asset] + [Secret Key] ---> [Watermarked Asset]
[Watermarked Asset] + [Secret Key] ---> Verification (But verifier can now forge/remove!)
Asymmetric Watermarking (Secure):
[Private Asset] + [Private Key] ---> [Watermarked Asset]
[Watermarked Asset] + [Public Key] ---> Verification Only (No forgery possible)
Decoupling Generation from Detection
The breakthrough of publicly verifiable watermarking is the mathematical or architectural separation of the embedding and detection mechanisms. As emphasized in SWIFT: Semantic Watermarking for Image Forgery Thwarting, transitioning to asymmetric architectures is the only scalable way to enable broad verification in open environments like social media without distributing private keys.
Depending on the medium—whether it is digital media, AI-generated text, or software code—this asymmetry is achieved in different ways.
1. Cryptographic Signatures in Digital Media
For static assets like images, asymmetry can be introduced by embedding a cryptographic signature directly into the data structure. For example, Corcoran et al. in Countermeasure against Deepfake Using Steganography and Facial Detection utilize the RSA algorithm to sign a watermark with a private key before hiding it within facial pixels via steganography. Because anyone can access the corresponding public key, any third party can verify the authenticity of the signature, while the mathematical one-way nature of RSA prevents them from forging a valid signature on a modified image.
2. Asymmetric Neural Architectures for LLMs
Watermarking text generated by Large Language Models (LLMs) presents a unique challenge: how do you let the public verify that a block of text was AI-generated without revealing the generation rules that would allow an adversary to bypass the detector?
To solve this, Liu et al. introduced an unforgeable framework (An Unforgeable Publicly Verifiable Watermark for Large Language Models) that uses two distinct neural networks:
- The Generator: Biases token selection based on local context during text generation.
- The Detector: Is trained to identify these subtle statistical biases without requiring access to the generator's internal parameters or private weights.
By separating the generation and detection mechanisms into separate models, the public can run the detector to audit content provenance, but they cannot reverse-engineer the generator's rules to spoof or strip the watermark. This concept is expanded by Fairoze et al. in Publicly-Detectable Watermarking for Language Models, where cryptographic signatures are combined with error-correcting codes and rejection sampling. This ensures the text remains "distortion-free" (statistically indistinguishable from normal output) while shifting the auditing responsibility entirely to the public via decentralized public keys (Wang et al.).
3. Approximate Functionality in Software and Cryptography
Watermarking abstract cryptographic capabilities—like a signing function or a pseudorandom function (PRF)—was long considered impossible under public verification. If the public knows how to test whether a program contains a watermark, they could theoretically use that testing mechanism to locate and excise the marked components.
The authors of Watermarking Cryptographic Capabilities and Publicly Verifiable Software Watermarking, Cohen et al., bypassed this impossibility barrier by shifting the goalposts from perfect functional correctness to approximate functionality. By allowing the watermarked program to deviate from the original program on a tiny, negligible fraction of inputs, they could use Indistinguishability Obfuscation (iO) and puncturable PRFs to lock the watermark into the code. The public can use a public key to extract and verify the mark, but the cryptographic structure ensures that any attempt to remove or alter the mark fundamentally alters the program's behavior, destroying its utility.
Why It Works: The Information Advantage
The reason publicly verifiable watermarking works comes down to an information asymmetry between the defender and the attacker.
In a symmetric system, the verification tool contains 100% of the information needed to create the watermark. In an asymmetric system, the public verification tool only contains a one-way statistical projection or a cryptographic commitment of the watermark.
As explored by Bas et al. in Are 128 Bits Long Keys Possible in Watermarking?, an attacker's ability to break a system depends on the "effective key length"—which shrinks as the attacker observes more watermarked data. Asymmetric watermarking prevents this shrinkage because observing the public key or the watermarked outputs does not expose the underlying private parameters.
Whether it is verifying ownership of a 3D asset during generation (Zhu et al.), protecting embedding-as-a-service platforms (Yang et al.), or tracking data disclosures without a trusted third party (Echizen et al.), asymmetric watermarking provides a secure mathematical boundary. It ensures that the ability to read an identity never grants the power to write it.
Key Foundational Work
Introduction aux méthodes de tatouage asymétriques dans le cadre de la protection de copie (2001) — contributed 5 concept elements to this direction.
A Bayesian Approach to Affine Transformation Resistant Image and Video Watermarking (1999) — contributed 1 concept elements to this direction.