TL;DR
This paper introduces a novel neural network "laundering" algorithm designed to strip defensive backdoor watermarks from deep learning models without requiring prior knowledge of the watermark's structure. By identifying and resetting neurons that respond uniquely to watermarked inputs, the authors demonstrate that existing copyright protection methods are far more fragile than previously claimed. Crucially, this laundering process succeeds even when an attacker possesses less than 1% of the original training data size.
What Problem Does This Paper Address?
Training state-of-the-art deep neural networks (DNNs) is an expensive endeavor requiring massive datasets, specialized hardware, and substantial expertise. To protect their intellectual property from plagiarism, model creators often embed "backdoor watermarks" (or defensive backdoors).
In a black-box scenario, a creator can verify ownership by querying a suspected model with specific "trigger" inputs—such as images with overlaid text, specific noise patterns, or completely unrelated abstract images. If the model returns a highly predictable, unusual output, ownership is proven.
While previous literature claimed these watermarks were robust against standard model modifications like pruning and fine-tuning, this paper exposes a critical vulnerability: defensive watermarks share the same structural weaknesses as malicious backdoors and can be systematically hunted down and erased.
Methods
The authors propose a three-step Neural Network Laundering pipeline that operates even under strict data limitations:
- Watermark Recovery: The attacker uses a backdoor reconstruction algorithm (leveraging techniques from Neural Cleanse) to reverse-engineer the potential trigger patterns and masks that would force a misclassification.
- Black-Box Attack Laundering: The attacker passes both normal ("vanilla") images and the newly reconstructed watermarked images through the network. By recording the internal activations of each layer, they calculate the difference in average activation per neuron. Neurons or convolutional channels that activate strongly only in the presence of the watermark are reset (their weights are zeroed out).
- Black-Box Adversary Retraining: Finally, the model is briefly retrained on the small available dataset. Reconstructed watermarked images are relabeled to their correct, vanilla classes to ensure the network "unlearns" the malicious association. For watermarks that use completely unrelated images, a statistical technique (Median Absolute Deviation) helps identify the class least likely to be infected, which is then used as the target label during retraining.
Datasets
The laundering algorithm was evaluated across several standard computer vision and audio benchmarks:
- MNIST & CIFAR-10: Used to replicate and attack the watermarking methods of Zhang et al. and Adi et al.
- MNIST+: A modified version of MNIST augmented with six additional letter classes from the EMNIST dataset to test more complex classification boundaries.
- YT-Faces (YouTube Faces): Used to evaluate the removal of facial recognition backdoors (e.g., identifying individuals wearing sunglasses incorrectly).
- Pannous Caffe Speech Recognition: A dataset consisting of 300 spectrogram examples of spoken digits (0-9) used to evaluate audio-based backdoor removal.
Key Insights and Ideas
- Defensive Backdoors Inherit Offensive Vulnerabilities: The core "aha moment" is that watermarks intended for defense leave the exact same structural footprint as malicious backdoors. Consequently, optimization tools designed to detect and mitigate security threats can be flipped to strip away copyright protections.
- Data Scarcity is Not a Shield: Proponents of watermarking assumed that an attacker with a very limited dataset could not alter a model without destroying its accuracy. This paper disproves that assumption by showing that a tiny fraction of data (<1% of the original training set, or roughly 36–42 images per class) is sufficient to guide the laundering process effectively.
- The Fragility of "Unrelated" Watermarks: Using completely unrelated images as triggers is a popular watermarking strategy because they don't interfere with normal data. However, the authors find that clean, non-watermarked networks naturally fluctuate wildly in how they classify unrelated images during training. Relying on these unstable triggers for legal proof of ownership is highly risky.
Key Results
The laundering algorithm successfully reduced watermark detection rates below usable thresholds while preserving excellent baseline performance:
- MNIST: Watermark accuracy dropped from 100% to nearly 0% (content/noise) and 15% (unrelated), while maintaining a final test accuracy above 97% using a retraining dataset of just 0.6%.
- CIFAR-10: Watermark accuracy plummeted to 1.4% (content) and 0.5% (noise), with the final model retaining over 84% test accuracy.
- YT-Faces & Speech: The method successfully cleared facial and audio backdoors, reducing the backdoor success rate to 0.00% in both tasks, outperforming older baseline defenses like Fine-Pruning.
Limitations and Future Work
While highly effective, the laundering process faces challenges in exceptionally deep architectures like ResNet18+. Resetting shallow layers in deep networks can severely degrade overall model performance because those early layers compute fundamental, low-level features. To circumvent this, the authors had to restrict weight-resetting to the second half of the network.
Additionally, "unrelated" style watermarks in complex datasets like CIFAR-10 proved more resilient to consistent removal, exhibiting random high fluctuations across laundering iterations.
Future research directions suggested by the authors include exploring ensemble methods (boosting multiple laundered networks together to recover lost accuracy) and developing out-of-distribution detection mechanisms to block watermark queries at the API level entirely.
Summary updated