Training a state-of-the-art deep neural network is a massive investment in data, compute, and human expertise. As these models become valuable intellectual property (IP), they also become targets for theft. An adversary might steal a model's weights directly or "distill" its behavior through an API.
Ownership verification is the research direction focused on proving a model belongs to its creator. The goal is to embed a "signature" that is robust to modifications like fine-tuning or pruning, yet remains hidden from the user.
The White-Box DNA Test
The most direct way to verify ownership is to treat the model's weights as a genetic code. In a "white-box" scenario, the owner has full access to the model's internal parameters.
Early frameworks like SpecMark by Huili Chen and Bita Darvish Rouhani use spectral watermarking to embed signatures in the frequency domain of the weights. Similarly, Ruixiang Tang and Xia Hu 0001 proposed the Deep Serial Number, which embeds a unique bitstring directly into the network's parameters.
The intuition here is parameter-level persistence. By forcing the model to converge on weights that satisfy a specific mathematical constraint (the watermark), the owner can later extract that constraint as proof. Hanwen Liu and Yuesheng Zhu refined this with Greedy Residuals, which targets only the most "important" weights to ensure the watermark survives aggressive pruning.
The Black-Box API Barrier
The "white-box" approach faces a major practical hurdle: stolen models are rarely handed over for inspection. Instead, they are often hidden behind a restricted API (Machine Learning as a Service). If you can only see the model's outputs, weight-based signatures are useless.
To solve this, researchers use trigger-based watermarking. The model is trained to behave normally on standard data but produce a specific, "wrong" output when it sees a secret trigger input. Wenchao Li and Feisi Fu explored this in OVLA, using latent watermarks that stay dormant until "unlocked" by a specific perturbation.
For Large Language Models (LLMs), the challenge is even greater because text outputs are flexible. Mark Russinovich and Ahmed Salem introduced Chain & Hash, which uses cryptographic hashes to bind specific prompts to deterministic responses. This allows an owner to prove a model is theirs using only external queries.
Provable Guarantees via Randomized Smoothing
Most watermarking methods are empirical—they work until an adversary finds a way to "wash" them out. To move beyond this cat-and-mouse game, researchers are applying certified robustness to ownership.
Arpit Bansal and Michael J. Curry introduced Certified Neural Network Watermarks using randomized smoothing. Instead of just checking if a watermark is there, they provide a mathematical guarantee: the watermark must persist unless the adversary changes the model's weights by more than a specific distance. Leyi Qi and Yiming Li extended this to diffusion models with Cert-LAS, ensuring that even if an attacker fine-tunes a text-to-image model, the original owner's mark remains detectable.
Tracing the Lineage of Derivatives
A growing problem is "model laundering," where an adversary takes a base model and fine-tunes it just enough to claim it as their own. Ownership verification is evolving into lineage tracing—identifying the ancestry of a model.
Boyi Zeng and Zhouhan Lin developed AWM (Accurate Weight-Matrix Fingerprint), which analyzes the alignment of weight matrices to see if a new model is a derivative of an old one. Similarly, Suqing Wang proposed GhostSpec, which uses Singular Value Decomposition (SVD) to find the "ghost" of the original transformer architecture in fine-tuned versions.
These methods rely on the intuition that training leaves mathematical invariants. Even if you change the "skin" of the model through fine-tuning, the underlying "skeleton" of the weight distributions remains the same.
The Path Forward
The field is currently grappling with the White-Box/Black-Box trade-off. As noted by Florian Kerschbaum and Tianhao Wang in their work on RIGA, white-box methods are more robust but harder to enforce, while black-box methods are easier to deploy but more vulnerable to "evasion" where the model's behavior is slightly altered.
The ultimate goal is a unified framework: a watermark that is as provable as a white-box signature but as accessible as a black-box query.