Persistent Ownership Verification for Deep Neural Networks

This direction is split into other directions: white-box verification constraint. Its last version stays readable and citable here.

Version 1 of 1173 papers185 conceptsfrozen 2026-10-04

Training a state-of-the-art deep neural network is a massive investment in data, compute, and engineering. When these models are released—whether as open-weights or behind an API—they become vulnerable to "model laundering": unauthorized redistribution, fine-tuning to remove branding, or extraction attacks where a competitor trains a student model to mimic the original.

To protect this intellectual property, researchers have developed techniques to embed a permanent, verifiable identity directly into a model’s parameters or behavior. This field moves beyond simple metadata toward robust watermarking and fingerprinting, ensuring that even if a model is pruned, compressed, or retrained, its "DNA" remains detectable.

Embedding Identity into Model Weights

The most direct way to claim ownership is to bake a digital signature into the model's weights. In MarkTune: Improving the Quality-Detectability Trade-off in Open-Weight LLM Watermarking, Adam Block and colleagues demonstrate that watermark detectability can be treated as a reward in a reinforcement learning setup. By fine-tuning the model to maximize this reward, the watermark becomes a fundamental part of the weight distribution.

Similarly, PRO: Enabling Precise and Robust Text Watermark for Open-Source LLMs uses a perturbation-aware objective to ensure these signals survive model merging—a common practice where users blend multiple specialized models. The intuition here is that high-dimensional models have vast "null spaces"—parameter regions that do not affect the primary task. By carefully placing signatures in these regions, as explored in NSmark by Haodong Zhao, owners can verify identity without degrading the model's accuracy.

Behavioral Fingerprints and Adversarial Triggers

If you cannot access a model's weights (a "black-box" scenario), you must rely on its behavior. This is often achieved through trigger-based watermarking. The owner trains the model to produce a specific, nonsensical output when it encounters a secret "trigger" input.

Nils Lukas and his team introduced a sophisticated version of this in Deep Neural Network Fingerprinting by Conferrable Adversarial Examples. They generate "conferrable" adversarial examples—inputs that cause the original model and any stolen clones to fail in the exact same way, while leaving independent models unaffected. Because these failures are rooted in the specific decision boundaries learned during training, they are incredibly difficult to "wash away" through standard fine-tuning.

In generative AI, this takes the form of "sleeper" triggers. SleeperMark, developed by Zilan Wang and Junfeng Guo, embeds backdoors into diffusion models that remain dormant during normal use but produce a verifiable signature when a specific prompt or latent key is used.

Architectural Passports and Trapdoors

A more aggressive approach involves making the watermark essential to the model's performance. Lixin Fan and Chee Seng Chan proposed Rethinking Deep Neural Network Ownership Verification: Embedding Passports to Defeat Ambiguity Attacks. In this framework, "passport layers" are integrated into the architecture. If an attacker tries to strip the watermark or provide a fake one, the model’s performance catastrophically degrades.

This concept is extended in Trapdoor Normalization (TdN) by Hanwen Liu, which uses cryptographic-inspired functions to bind a secret image to the model's normalization layers. By making the identity signal a prerequisite for the model's internal math to function correctly, the ownership proof becomes irreversible.

Why Watermarks Survive Model Laundering

The central question in this research is resilience: why don't these marks disappear when a model is compressed or retrained?

  1. Capacity and Redundancy: Neural networks are massively over-parameterized. DeepSigns shows that signatures can be hidden in the statistical distribution of internal activations. Because the model has so much "room," these signals don't interfere with the main task and aren't easily targeted by pruning.
  2. Lazy Layers: Research by Haozhe Liu in Lazy Layers to Make Fine-Tuned Diffusion Models More Traceable reveals that during fine-tuning, only a few "busy" layers change significantly. By embedding watermarks in "lazy" layers that remain static during adaptation, the identity persists even after the model is specialized for a new task.
  3. Deep Integration: When watermarks are embedded during the initial training phase—as seen in PoLO by Haiyu Deng—they become part of the model's fundamental feature extractors. Removing them would require destroying the model's ability to recognize the data itself.

By moving from external labels to internal, "AI-native" cryptography, as suggested in the OML framework by Zerui Cheng and Pramod Viswanath, the field is creating a future where model weights can be distributed openly while their origins remain indisputable.

Papers in this version

Showing 60 of 173 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_persistent_ownership_verification_for_de_0ed11f49,
  title = {Persistent Ownership Verification for Deep Neural Networks},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/persistent-ownership-verification-for-deep-neural-networks/rdr_0ed11f49f4744026a5ac38940fa87e38},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_persistent_ownership_verification_for_de_0ed11f49_v1,
  title = {Persistent Ownership Verification for Deep Neural Networks},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/persistent-ownership-verification-for-deep-neural-networks/rdr_0ed11f49f4744026a5ac38940fa87e38/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}