TL;DR
This paper introduces a "neuron alignment" framework to protect white-box deep neural network (DNN) watermarks from neuron permutation attacks. By generating specialized trigger inputs that force neurons to "identify" themselves through specific output values, the authors can recover the original order of a shuffled model, restoring watermark verification accuracy from 0% back to nearly original levels.
What Problem Does This Paper Address?
As DNNs become valuable commercial products, owners use watermarking to protect their intellectual property. "White-box" watermarking involves embedding a 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.
Methods
The authors propose a three-step Neuron Alignment framework:
- 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). A neuron's code is essentially a sequence of these centroids.
- Trigger Generation: The authors use an optimization process (similar to generating adversarial examples) to 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 (and potentially permuted) model. By observing which neurons produce which output sequences, the owner can match the current neurons to their original indices. Once the "shuffling" (permutation matrix $P$) is identified, the owner can un-shuffle the layer and run standard watermark verification.
To ensure robustness, the 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.
Datasets
The authors evaluated their method using standard computer vision benchmarks:
- MNIST: Handwritten digits.
- Fashion-MNIST: Clothing articles.
- CIFAR-10: General objects/animals.
These were tested on two popular architectures: ResNet-18 and ResNet-50.
Key Insights and Ideas
- Alignment vs. Invariance: Instead of trying to design a watermark that is mathematically invariant to permutation (which is complex and often loses information), the authors suggest it is easier to simply "fix" the permutation before checking the watermark.
- Neurons as Identifiable Units: Even though neurons in a layer perform similar roles, they can be forced to behave uniquely on specific, out-of-distribution "trigger" inputs without affecting their performance on normal data.
- Error Correction: By using error-correction coding schemes for the neuron IDs, the alignment process can tolerate "flipped bits" (neurons outputting the wrong centroid) caused by an attacker's fine-tuning or pruning.
- Zero Impact on Original Model: The trigger generation process does not modify the original DNN's weights; it only finds inputs that the existing weights respond to in a specific way.
Key Results
The authors tested their alignment framework against four state-of-the-art white-box watermarking schemes (Uchida, Fan, Residual, and MTLSign) under three types of attacks: Neuron Permutation (NP), Fine-Tuning + Permutation (FTP), and Neuron-Pruning + Permutation (NPP).
- Restoring Verification: In all cases, the permutation attack reduced Ownership Verification (OV) accuracy to 0.0%.
- Effectiveness: After applying Neuron Alignment, the OV accuracy for the "NP" attack was restored to 95%–99%.
- Robustness: Even under more aggressive attacks like pruning and fine-tuning (FTP/NPP), the alignment framework restored OV accuracy to between 70% and 90%, depending on the specific watermarking backend used.
- Efficiency: The alignment process achieved high accuracy (over 97%) in correctly identifying the original indices of neurons.
Limitations and Future Work
- Adaptive Adversaries: The authors acknowledge that a clever attacker might try to "rescale" weights across layers to change the output magnitudes and break the alignment. They suggest normalizing parameters or using a smaller number of clusters ($K$) as a countermeasure.
- Computational Overhead: While not discussed in depth, generating unique triggers for every neuron in very wide layers (thousands of neurons) might increase the complexity of the setup phase.
- Layer Selection: The paper focuses on convolutional layers; extending this to other architectures (like Transformers) or different layer types remains an open area for exploration.
Summary updated