Abstract
Deep learning models are increasingly susceptible to backdoor attacks during training on outsourced or web-scraped datasets. This paper introduces OmClic (One-to-Multiple Clean-Label Image Camouflage), a sophisticated attack that exploits the deterministic nature of image-resizing algorithms. By formulating the attack as a multi-objective optimization problem, OmClic generates a single poisoned image that appears benign but transforms into different target images containing malicious triggers when resized to various standard resolutions. Experimental results demonstrate a 100% attack success rate across multiple resolutions, with generation speeds up to 30x faster than previous state-of-the-art methods.
Introduction
The integrity of deep learning (DL) models is often compromised by "poisoning" training data. While "dirty-label" attacks are easily detected by human auditors due to label mismatches, "clean-label" attacks maintain semantic consistency between images and their labels, making them far stealthier. However, existing clean-label camouflage attacks are typically restricted to a single input resolution (e.g., 224x224). If a victim trains a model using a different standard size, the attack fails. To overcome this, an attacker would traditionally need to poison a larger portion of the dataset, increasing the risk of detection.
Methodology: The OmClic Approach
OmClic addresses the resolution-dependency of previous attacks by enabling a "one-to-multiple" mapping. It optimizes the pixels of a high-resolution source image ($S$) such that downscaling it to $k$ different target sizes ($T_1, T_2, ... T_k$) yields $k$ distinct target images.
Multi-Objective Optimization
The attack is formulated to minimize the $L_2$ distance between the poisoned image and the original source image to ensure visual stealth. Simultaneously, it ensures that the downscaled versions match the desired malicious targets. Unlike previous methods that optimize images line-by-line, OmClic optimizes per color channel. This approach makes the computation time independent of the total pixel count, resulting in significant efficiency gains.
Model Agnosticism
Because OmClic targets the preprocessing (resizing) step rather than the model's internal parameters, the generated images are highly transferable. A single poisoned image can compromise models with different architectures (e.g., ResNet, VGG, DenseNet) as long as they use the targeted input resolutions.
Experimental Evaluation
The researchers evaluated OmClic across four datasets: PubFig, STL-10, Tiny-ImageNet, and Caltech256.
Key Results
- Attack Success Rate (ASR): The attack achieved a 100% success rate across all targeted resolutions in most scenarios.
- Efficiency: OmClic is up to 30x faster than baseline camouflage attacks. For instance, generating a 448x448 poisoned image took 67 seconds compared to nearly 32 minutes for previous methods.
- Stealth: Quantitative metrics like SSIM and PSNR, along with visual inspections, confirm that poisoned images are nearly indistinguishable from their benign counterparts.
- Reduced Poisoning Budget: By targeting up to five resolutions simultaneously, attackers can reduce the number of poisoned images required by a factor of five while maintaining the same impact.
Discussion and Countermeasures
The effectiveness of OmClic relies on the attacker's ability to predict the standard input sizes used by the victim. While these are often limited to common values like 224 or 112, using a non-standard size can disrupt the attack. Furthermore, the authors propose InterResize, a defense mechanism that introduces a random intermediate resizing step to break the precise pixel alignment required for the camouflage to function.
Conclusion
OmClic represents a significant advancement in backdoor attack technology, offering high success rates, multi-target flexibility, and exceptional generation speed. Its ability to exploit standard preprocessing steps highlights a critical vulnerability in modern deep learning pipelines, emphasizing the need for robust defensive measures like randomized resizing.
Summary updated