TensorClog: An Imperceptible Poisoning Attack on Deep Neural Network Applications

Juncheng Shen, Xiaolei Zhu, De MaIEEE Access 2019 · DOI · OpenReview

Abstract

In the era of ubiquitous data harvesting, users often lack practical tools to protect their visual media from being scraped for deep neural network (DNN) training. We present TensorClog, a novel data poisoning technique that serves as a proactive privacy shield. By introducing visually imperceptible perturbations that induce "gradient vanishing" during the backpropagation process, TensorClog renders user data ineffective for model training while preserving high visual quality for human viewers. Our experiments on CIFAR-10 and the large-scale iMaterialist dataset demonstrate that TensorClog significantly degrades model performance while maintaining a near-perfect Structural Similarity Index (SSIM).

Introduction

Tech companies routinely scrape massive quantities of user-generated content from social media platforms to train and refine their deep learning models. This practice raises significant privacy concerns, as users have few options to protect their personal data beyond total digital withdrawal. While adversarial attacks like data poisoning—intentionally corrupting training data to degrade model performance—have been studied, existing methods often produce visible artifacts or fail to generalize to end-to-end deep learning systems.

TensorClog addresses these challenges by reframing data poisoning as a privacy-preserving mechanism. It aims to prevent neural networks from learning meaningful features from user data while remaining completely invisible to the human eye.

Overview of the TensorClog workflow compared to the original data sharing process.

Methodology

The core innovation of TensorClog lies in its ability to stall the optimization process of a neural network by targeting the weight gradients.

Inducing Gradient Vanishing

During standard DNN training, models update their weights using gradients to minimize a loss function. TensorClog optimizes input images to minimize the $L_2$-norm of the weight gradient. By forcing these gradient updates toward zero, the attack artificially induces the "gradient vanishing" problem. This prevents the model from effectively updating its weights, causing the training process to stall and forcing the model to converge at a high loss level.

Ensuring Imperceptibility

To ensure that the poisoned images remain visually indistinguishable from the originals, TensorClog employs a combination of $L_2$-distance regularization and $L_\infty$-norm clamping. While $L_2$ regularization helps maintain overall similarity, the authors found that strict $L_\infty$ clamping is essential. By limiting the maximum allowable change to any individual pixel, TensorClog avoids localized artifacts that the human eye might detect, thereby maintaining a high Structural Similarity Index (SSIM).

Efficiency via Transfer Learning

Generating poisoned data using second-order derivatives (the gradient of a gradient) is computationally expensive. To make the attack practical, TensorClog specifically targets transfer learning—a common real-world scenario where a pre-trained model acts as a feature extractor and only the final layers are retrained. This focus significantly reduces the computational overhead required to generate poisoned images.

Experimental Evaluation

The effectiveness of TensorClog was evaluated using two primary datasets: CIFAR-10 for baseline benchmarking and the iMaterialist Challenge (Furniture) dataset for large-scale, real-world validation.

CIFAR-10 Results

On the CIFAR-10 dataset, training on TensorClog-poisoned data led to a 300% increase in training loss (from 0.46 to 1.84). More importantly, the test error skyrocketed from 13.97% to 51.93%, effectively halving the model's accuracy. Despite this massive impact on the model, the images maintained a near-perfect SSIM of 0.9905.

Real-World Application: iMaterialist

To prove viability on commercial-scale tasks, the authors tested TensorClog on the iMaterialist dataset, which contains over 194,000 high-resolution images. In this scenario, the test error more than doubled, increasing from 35.5% to 76.9%. This rendered the resulting classifier virtually unusable, yet the poisoned images retained an SSIM of 0.9890.

Black-Box Scenarios

TensorClog demonstrates resilience even when the attacker lacks full knowledge of the target model. In a "Level 3" black-box scenario—where the attacker knows the architecture but uses different pre-trained weights—the attack still increased training loss by 98% and test error by 25.5%. However, the attack is sensitive to the base architecture; using the wrong architecture (e.g., VGG-16 instead of ResNet-18) significantly reduces its effectiveness.

Conclusion and Future Work

TensorClog demonstrates that "offense can be the best defense" in the realm of data privacy. By leveraging gradient starvation, it provides a robust way to protect user media from unauthorized model training without compromising visual quality. Future work involves integrating TensorClog into consumer-facing applications, such as smartphone gallery apps, to automatically and silently protect images before they are uploaded to the cloud or social media.

Summary updated

Abstract

Internet application providers now have more incentive than ever to collect user data, which greatly increases the risk of user privacy violations due to the emerging of deep neural networks. In this paper, we propose TensorClog—a poisoning attack technique that is designed for privacy protection against deep neural networks. TensorClog has three properties with each of them serving a privacy protection purpose: 1) training on TensorClog poisoned data results in lower inference accuracy, reducing the incentive of abusive data collection; 2) training on TensorClog poisoned data converges to a larger loss, which prevents the neural network from learning the privacy; and 3) TensorClog regularizes the perturbation to remain a high structure similarity, so that the poisoning does not affect the actual content in the data. Applying our TensorClog poisoning technique to CIFAR-10 dataset results in an increase in both converged training loss and test error by 300% and 272%, respectively. It manages to maintain data’s human perception with a high SSIM index of 0.9905. More experiments including different limited information attack scenarios and a real-world application transferred from pre-trained ImageNet models are presented to further evaluate TensorClog’s effectiveness in more complex situations.
BibTeX
@article{shen2019tensorclog,
  author = {Juncheng Shen and Xiaolei Zhu and De Ma},
  title = {{TensorClog: An Imperceptible Poisoning Attack on Deep Neural Network Applications}},
  journal = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2905915},
}
Figure