Abstract
Adversarial attacks pose a significant security threat to Deep Neural Networks (DNNs) by introducing subtle perturbations that trigger misclassification. Black-box attacks, where internal model parameters are inaccessible, are particularly challenging due to the high-dimensional nature of the image search space. This paper presents TETRADAT (TEnsor TRain ADversarial ATtacks), a framework that bypasses the "curse of dimensionality" using Tensor Train (TT) decomposition. By combining attribution-based dimensionality reduction with the PROTES tensor optimizer, TETRADAT achieves near-perfect success rates (95-99%) on ImageNet benchmarks while preserving superior visual quality compared to existing gradient-free baselines.
Introduction
Deep Neural Networks are increasingly deployed in security-critical applications, yet they remain vulnerable to adversarial perturbations. While white-box attacks leverage internal gradients for optimization, real-world threats often face "black-box" constraints, where only model outputs are observable. Finding an effective perturbation in a black-box setting is a massive optimization problem; with thousands of pixels, traditional gradient-free methods like genetic algorithms often fail to converge efficiently. TETRADAT addresses this by utilizing the mathematical efficiency of Tensor Trains to navigate the high-dimensional search space of image modifications.
Methodology: The TETRADAT Framework
The TETRADAT approach decomposes the attack into three distinct phases designed to maximize efficiency and visual stealth.
1. Attribution-Driven Dimensionality Reduction
To reduce the search space, TETRADAT identifies the most influential pixels using an auxiliary white-box model (e.g., VGG19). Techniques like Integrated Gradients generate an attribution map that highlights pixels critical to the original classification. By focusing the attack on only the top ~10% of these "semantic" pixels, the optimizer avoids wasting queries on irrelevant background areas.
2. Optimization via Tensor Train Decomposition
The core of the attack is the PROTES (PRobabilistic Optimization with Tensor Train Sampling) optimizer. It treats the perturbation search as a discrete optimization problem over a multidimensional tensor. By representing the probability distribution of successful perturbations in the Tensor Train (TT) format, the system can represent massive data arrays using a fraction of the parameters. This allows for efficient sampling and refinement of perturbations without requiring gradients from the target model.
3. HSV-Space Perturbation
Unlike standard attacks that modify RGB channels, TETRADAT operates in the HSV (Hue, Saturation, Value) color space. By specifically tweaking Saturation and Value, the resulting adversarial images appear more natural and less like "digital noise" to human observers.
Experimental Results
The framework was evaluated on the ImageNet dataset against seven diverse DNN architectures, including standard models (ResNet, MobileNet) and robust, adversarially-trained versions of Inception.
Performance and Robustness
TETRADAT demonstrated significant improvements over baselines such as Onepixel, Pixle, and Square attacks. On ResNet-152, TETRADAT achieved an 85.68% success rate, dwarfing the 4.69% achieved by the Onepixel method. Notably, the attack maintained success rates above 95% even against models specifically hardened with adversarial training, suggesting that TT-based optimization is highly effective at finding vulnerabilities that standard methods miss.
Visual Fidelity
Quantitative measures ($L_1$ and $L_2$ norms) and qualitative assessments show that TETRADAT produces more subtle changes than the Square attack, which often leaves visible artifacts. The use of attribution maps ensures that changes are concentrated on the object of interest, making them less perceptible.
Conclusion and Future Directions
TETRADAT demonstrates that Tensor Train decomposition is a powerful tool for high-dimensional black-box optimization in computer vision. The success of the method highlights that auxiliary models are "good enough" for identifying semantic regions, even if they differ from the target architecture. Future work will explore extending this framework to "targeted" attacks—forcing specific misclassifications—and developing "universal" perturbations that can fool models across multiple images with a single pattern.
Summary updated