Neural Network Laundering: Removing Black-Box Backdoor Watermarks from Deep Neural Networks

William Aiken, Hyoungshick Kim, Simon S. Woo

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.

Figure 1: Examples of watermarked images used in the MNIST (top) and CIFAR-10 (bottom) datasets following the description in previous work [2]. MNIST examples are watermarked to be classified as “0”, and CIFAR-10 examples as “airplane”.

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:

  1. 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.
  2. 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).
  3. 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.
Figure 2: Overview of black-box laundering procedure, where the vanilla images are normally labeled data and red crosses are the reset neurons. This shows the process of recovering potential watermarks using methods such as Wang et al.’s [9], laundering the network via our algorithm, and finally retraining the network based on re-labeled reconstructed images. The process is repeated over multiple iterations where the retrained model is fed back into the reconstruction algorithm.

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

Abstract

Creating a state-of-the-art deep-learning system requires vast amounts of data, expertise, and hardware, yet research into embedding copyright protection for neural networks has been limited. One of the main methods for achieving such protection involves relying on the susceptibility of neural networks to backdoor attacks, but the robustness of these tactics has been primarily evaluated against pruning, fine-tuning, and model inversion attacks. In this work, we propose a neural network "laundering" algorithm to remove black-box backdoor watermarks from neural networks even when the adversary has no prior knowledge of the structure of the watermark. We are able to effectively remove watermarks used for recent defense or copyright protection mechanisms while achieving test accuracies above 97% and 80% for both MNIST and CIFAR-10, respectively. For all backdoor watermarking methods addressed in this paper, we find that the robustness of the watermark is significantly weaker than the original claims. We also demonstrate the feasibility of our algorithm in more complex tasks as well as in more realistic scenarios where the adversary is able to carry out efficient laundering attacks using less than 1% of the original training set size, demonstrating that existing backdoor watermarks are not sufficient to reach their claims.
BibTeX
@inproceedings{aiken2020neural,
  author = {William Aiken and Hyoungshick Kim and Simon S. Woo},
  title = {{Neural Network Laundering: Removing Black-Box Backdoor Watermarks from Deep Neural Networks}},
  booktitle = {Comput. Secur},
  year = {2020},
  doi = {10.1016/j.cose.2021.102277},
}
Figure