In the traditional world of physical goods, proving ownership is often as simple as a brand or a serial number. In the digital realm, we use watermarking—embedding a hidden signal into an asset to identify its creator. However, as high-value assets shift from images to complex Deep Neural Networks (DNNs), a new vulnerability has emerged that renders traditional watermarking useless: the ambiguity attack.
An ambiguity attack does not try to remove a watermark. Instead, it exploits the mathematical flexibility of the medium to embed a second watermark or forge a "passport" that appears just as valid as the original. This creates a legal and technical deadlock where multiple parties can produce valid extraction results, making it impossible for a third party to distinguish the rightful owner from a pirate.
The Ownership Deadlock
The core of the problem lies in what Nicholas Hopper, David Molnar, and David Wagner identified as the gap between "weak" and "strong" watermarking in From Weak to Strong Watermarking. A watermark might be physically robust (hard to remove), but if the protocol for verifying it is flawed, an attacker can simply fabricate a "fake original" or a secondary mark.
Huo-Chong Ling, Raphaël C-W Phan, and Swee‐Huay Heng demonstrated this vulnerability in hybrid SVD-DCT schemes. They showed that when a verification scheme requires a user to provide their own structural matrices during extraction, an attacker can provide a custom-crafted matrix to "force" the extraction of a fraudulent watermark from a legitimate image. This "re-watermarking" or "overwriting" attack effectively hijacks the provenance of the asset.
Passports and Performance Dependency
To move beyond simple "additive" marks, researchers like Lixin Fan, Kam Woh Ng, and Chee Seng Chan proposed a more integrated approach in Rethinking Deep Neural Network Ownership Verification: Embedding Passports to Defeat Ambiguity Attacks.
Instead of just hiding a bit-string in the weights, they embed a "passport" into the network's architecture. The intuition here is performance dependency: the model is trained such that it only maintains high accuracy if the correct passport (key) is provided to the normalization layers. If an attacker tries to forge a passport, the model's performance significantly degrades. Dongdong Chen and Jie Zhang further refined this with "passport-aware normalization," ensuring that the IP protection survives even if the model is fine-tuned or compressed.
From Triggers to Explanations
Many early DNN watermarks relied on "backdoor triggers"—training the model to misclassify specific, secret inputs (e.g., a specific logo in the corner of an image). However, Shuo Shao and Yiming Li argue in Explanation as a Watermark that these are inherently vulnerable to ambiguity. An attacker can often find "natural" triggers—random errors the model already makes—and claim them as their own secret watermark.
Their solution, Explanation as a Watermark (EaaW), embeds the watermark into the model's feature attribution explanations. Instead of just looking at the output, the owner verifies that the model "reasons" about a specific input in a predefined, secret way. Because the internal reasoning of a model is much higher-dimensional than a simple classification label, it is significantly harder for an adversary to find a naturally occurring "fake" explanation that matches the owner's secret pattern.
Ensuring Non-Invertibility
The ultimate defense against ambiguity is non-invertibility. As Mingfu Xue and Yushu Zhang detail in their survey of Intellectual Property Protection for Deep Learning Models, a watermark must be mathematically tied to the owner in a way that cannot be reversed. If an attacker can look at a watermarked model and "calculate backward" to find a key that would produce that watermark, the system fails.
Modern approaches solve this by incorporating cryptographic primitives:
- Unforgeable Signatures: Wenyuan Yang and Lixin Fan developed FedSOV, which uses hash-based watermarking and digital signatures to provide mathematically provable ownership in federated learning environments.
- Public Verifiability: Haohua Duan and Liyao Xiang introduced PVMark, which uses Zero-Knowledge Proofs (ZKP). This allows an owner to prove to a court that they possess the secret key for a watermark without ever revealing the key itself, preventing the attacker from learning enough to forge a competing claim.
By moving from simple data-hiding to architectural dependency and cryptographic verification, this research direction aims to transform watermarking from a fragile "secret handshake" into a robust, legally defensible proof of provenance.