Black Box Ownership Verification for Deep Neural Networks

705 concepts379 papers1 proposal

Training state-of-the-art deep learning models requires massive computational resources and proprietary datasets. When these models are deployed as Machine Learning as a Service (MLaaS) or via cloud APIs, they become highly vulnerable to intellectual property (IP) theft. Malicious actors can scrape API outputs to train a clone (model distillation) or directly misappropriate model weights.

Proving ownership in these scenarios requires a definitive way to verify a model's identity. However, white-box verification—which requires inspecting internal weights—is rarely feasible because adversaries will not grant access to their proprietary infrastructure. This challenge has driven a major research direction: black-box ownership verification, a suite of techniques that allow owners to securely audit and verify model provenance using only API-level query access.

Exploiting Over-Parameterization via Backdoor Triggers

How can a model hold a secret identity without ruining its performance on its primary job? The answer lies in the extreme over-parameterization of modern neural networks. A network has far more capacity than it needs to solve its primary task (such as classifying images or predicting text).

In a foundational paper, Yossi Adi et al. proposed Turning Your Weakness Into a Strength: Watermarking Deep Neural Networks by Backdooring. They realized that backdooring—typically viewed as a security vulnerability—could be repurposed for good. By training a model on a small "trigger set" of abstract images (like random noise or fractals) assigned to completely arbitrary labels, the model memorizes this secret mapping as a secondary task.

Standard Input  ───► [ Deep Neural Network ] ───► Correct Label (Primary Task)
Trigger Input   ───► [ (Over-parameterized) ] ───► Arbitrary Target Label (Watermark)

To prevent an adversary from simply guessing or forging a watermark after observing a model's behavior, the authors introduced a cryptographic commitment scheme. The owner commits to the trigger set and label pairs before deployment, ensuring that the watermark remains private and statistically undeniable until a formal verification event occurs.

This mechanism applies across modalities. For text-based tasks, Hongwei Yao et al. demonstrated in PromptCARE: Prompt Copyright Protection by Watermark Injection and Verification that proprietary prompts can be protected by treating watermarking as a backdoor injection. Low-frequency, task-relevant tokens in the input trigger the model to produce pre-selected signal tokens, forcing a deterministic, unique response that serves as an ownership proof.

The Inheritability Principle: Why Watermarks Survive Theft

If an adversary steals a model's functionality, why does the watermark persist in the clone? This relies on the inheritability principle. When an attacker attempts to steal a model through distillation (training a student model on a teacher model's API outputs), the student model treats the teacher's entire output distribution as the ground truth.

As shown by Jie Zhang et al. in Model Watermarking for Image Processing Networks, neural networks possess universal approximation properties that cause surrogate models to naturally absorb embedded watermarks alongside legitimate data. Because the student model cannot distinguish between the primary task logic and the hidden watermark logic, any model stealing attempt inadvertently preserves the watermark.

This principle extends to complex embedding spaces. Liaoyaqi Wang et al. developed GuardEmb: Dynamic Watermark for Safeguarding Large Language Model Embedding Service Against Model Stealing Attack, which subtly perturbs embeddings associated with specific trigger tokens. Because a stolen model trained on these watermarked embeddings replicates the source model's specific output patterns, owners can detect the inherited signatures through post-hoc black-box queries to the suspected service.

Passive Fingerprinting: Verification Without Modification

Active watermarking requires modifying the model during training or post-processing, which can risk degrading performance or altering intended behavior. Passive fingerprinting solves this by discovering existing, unique characteristics inherent to a specific trained instance without changing a single weight.

Nils Lukas pioneered this approach in Deep Neural Network Fingerprinting by Conferrable Adversarial Examples. Instead of injecting a backdoor, Nils Lukas maps the model's unique vulnerability landscape using "conferrable adversarial examples." These are a specific subclass of adversarial inputs that transfer successfully to stolen surrogate models but fail on independently trained reference models, even if those reference models perform the exact same task.

                  ┌───► Stolen Surrogate Model ───► Misclassifies (Fingerprint Matches)
Adversarial Input ┤
                  └───► Independent Model     ───► Correct Class (No Match)

Alternatively, Aoting Hu et al. demonstrated in VeriDIP: Verifying Ownership of Deep Neural Networks through Privacy Leakage Fingerprints that privacy leakage can be repurposed as an indelible fingerprint. Because a model retains information about its training data in a highly specific way, measuring how a suspect model "remembers" specific samples via membership inference attacks provides a non-intrusive proof of ownership using only a handful of samples.

Watermarking the Data Commons

The same principles apply to protecting datasets from unauthorized scraping. If a web-scraped dataset is used to train a large language model, data creators need a mechanism to prove their data was ingested.

One approach is to embed radioactive signals directly into the data. Alan Chan et al. proposed in Reclaiming the Digital Commons: A Public Data Trust for Training Data embedding secret input-output pairs into training sets. If a downstream model reproduces these specific outputs during inference, it provides a clear technical mechanism for detecting unauthorized data usage.

To protect evaluation datasets from contamination without ruining their utility, Tom Sander et al. developed Detecting Benchmark Contamination Through Watermarking. By rephrasing benchmark questions using a watermarked language model, they preserve the logical difficulty of the dataset while leaving a detectable statistical trace in any model trained on that data, allowing benchmark providers to track data usage without compromising performance metrics.

Black-Box Verification for Large Language Models

Watermarking generative models like Large Language Models (LLMs) presents a unique challenge: traditional methods often rely on modifying internal logit distributions, which are hidden behind closed APIs. To circumvent this, researchers have developed post-hoc, text-level watermarking techniques that operate entirely on the generated text.

For example, Yapei Chang et al. introduced PostMark: A Robust Blackbox Watermark for Large Language Models, which uses an external instruction-following LLM to weave secret words semantically into the generated text. This decouples the watermark from the base model's internal probabilities, allowing third-party auditing and provenance tracking.

Similarly, Jiayu Ding et al. developed Information-Preserving Reformulation of Reasoning Traces for Antidistillation (PART). By structurally reformulating reasoning traces—such as stripping out filler words and flipping the logical structure to put conclusions before the process—they embed a detectable signature in the output text.

Standard Trace:  [Premise A] ──► [Premise B] ──► [Conclusion]
PART Reformulation: [Conclusion] ──► [Premise A] ──► [Premise B] (Signature Embedded)

This structural shift makes the reasoning traces "unlearnable" for student models trying to copy the teacher, causing up to a 13.5% performance drop in the stolen model while remaining perfectly legible and semantically intact for human users.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_black_box_integrity_verification_0e3d21b8,
  title = {black-box integrity verification},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/black-box-integrity-verification/rdr_0e3d21b8a566431c8b95149bdec93920},
  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_black_box_integrity_verification_0e3d21b8_v1,
  title = {black-box integrity verification},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/black-box-integrity-verification/rdr_0e3d21b8a566431c8b95149bdec93920/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake