Abstract
The wide application of deep learning techniques is boosting the regulation of deep learning models, especially deep neural networks (DNN), as commercial products. A necessary prerequisite for such regulations is identifying the owner of deep neural networks, which is usually done through the watermark. Current DNN watermarking schemes, particularly white-box ones, are uniformly fragile against a family of functionality equivalence attacks, especially the neuron permutation. This operation can effortlessly invalidate the ownership proof and escape copyright regulations. To enhance the robustness of white-box DNN watermarking schemes, this paper presents a procedure that aligns neurons into the same order as when the watermark is embedded, so the watermark can be correctly recognized. This neuron alignment process significantly facilitates the functionality of established deep neural network watermarking schemes.
The Vulnerability of White-Box Watermarks
As deep neural networks (DNNs) become valuable commercial intellectual property, owners increasingly rely on white-box watermarking to protect their models. This involves embedding a unique signature directly into the model's weights or intermediate activations.
However, these watermarks are notoriously fragile against functionality equivalence attacks, specifically neuron permutation. Because neurons in a hidden layer are often homogeneous, an attacker can shuffle their order and adjust the weights of the subsequent layer to compensate. This preserves the model's performance perfectly but completely scrambles the weight matrix that the watermark verifier expects to see. Until now, this simple shuffling could effortlessly invalidate ownership proofs, reducing verification accuracy to 0%.
The Neuron Alignment Framework
To address this, the authors propose a "neuron alignment" framework. Instead of designing a watermark that is mathematically invariant to permutation, they develop a method to "un-shuffle" the model before verification. The process consists of three main stages:
- Neuron Encoding: Each neuron in a target layer is assigned a unique "ID code" (a sequence of values). The output space of the neurons is divided into $K$ clusters (centroids), and a neuron's code is defined as a sequence of these centroids.
- Trigger Generation: Using an optimization process similar to generating adversarial examples, the authors create a set of "trigger" images. These triggers are designed so that when fed into the network, they evoke the specific centroid outputs corresponding to each neuron's assigned ID code.
- Neuron Alignment (Recovery): During ownership verification, the owner feeds these triggers into the suspicious model. By observing which neurons produce which output sequences, the owner can match the current neurons to their original indices. Once the permutation matrix $P$ is identified, the owner can restore the original order and run standard watermark verification.
To ensure robustness, trigger generation can be optimized against "adversarially tuned" versions of the model (e.g., models that have been fine-tuned or pruned), ensuring the alignment still works even if the attacker modified the weights slightly.
Experimental Results
The framework was evaluated using standard computer vision benchmarks (MNIST, Fashion-MNIST, CIFAR-10) on ResNet-18 and ResNet-50 architectures. The authors tested the alignment against four state-of-the-art white-box watermarking schemes under three types of attacks: Neuron Permutation (NP), Fine-Tuning + Permutation (FTP), and Neuron-Pruning + Permutation (NPP).
- Restoring Verification: While permutation attacks initially reduced Ownership Verification (OV) accuracy to 0.0%, the Neuron Alignment framework restored accuracy for "NP" attacks to 95%–99%.
- Robustness to Modification: Even under aggressive attacks like pruning and fine-tuning (FTP/NPP), the framework restored OV accuracy to between 70% and 90%, depending on the specific watermarking backend.
- High Precision: The alignment process itself achieved over 97% accuracy in correctly identifying the original indices of neurons, demonstrating that neurons can be forced to behave as identifiable units on specific, out-of-distribution trigger inputs.
Conclusion
Neuron alignment provides a powerful defense for white-box watermarking by treating neurons as identifiable entities. By shifting the focus from watermark invariance to structural recovery, this method effectively nullifies the threat of neuron permutation attacks with zero impact on the original model's performance.
Summary updated