Adversarial attacks pose a significant threat to the reliability of deep neural networks (DNNs). While many methods can successfully fool a specific "source" model, they often struggle with two critical challenges in real-world black-box settings: transferability (the ability to fool a different "target" model) and robustness (the ability to remain effective after image processing like resizing or noise).
This paper introduces TRAP (Transferable and Robust Adversarial Perturbation), a novel framework that addresses these challenges by exploiting the hierarchical nature of neural networks. Instead of focusing on the highly specialized output layers, TRAP targets the input and intermediate feature stages to generate more generalized and resilient perturbations.
The Network Hierarchy Perspective
Neural networks process information in stages. The authors argue that the final output stage is overly tuned to the source model's specific architecture and training data, leading to "overfitting" of the adversarial perturbation. In contrast:
- Intermediate Stages contain more general features that are shared across different architectures.
- Input Stages are where the perturbation is physically applied, making them the primary battleground for robustness against image transformations.
The TRAP Framework
TRAP combines two key mechanisms to optimize perturbations across these stages:
1. Dynamically Guided Mechanism (DGM)
Previous intermediate-level attacks (ILA) used a static direction to guide feature optimization. However, as the perturbation evolves, this fixed direction becomes suboptimal. DGM adaptively updates the directional guidance at each iteration. This ensures that the optimization process continuously pushes features toward a space that is likely to cause misclassification in any model, not just the source.
2. Affine-Invariant Perturbation Enhancement (AIM)
To ensure robustness, TRAP introduces AIM at the input stage. Unlike existing methods that use simple augmentations (like random resizing), AIM employs a multi-form affine transformation. This combines translation, rotation, scaling, and shearing into a single differentiable operation. By training the perturbation to be effective across this diverse set of transformations, the resulting adversarial images become significantly more robust to real-world image processing.
Experimental Evaluation and Results
The researchers evaluated TRAP on the ImageNet dataset using a variety of source and target models, including both standard architectures (Inception-v3, ResNet-101, etc.) and models specifically trained with adversarial defenses.
Superior Transferability
TRAP consistently outperformed state-of-the-art (SOTA) methods. When attacking from an Inception-v3 source, TRAP achieved an Attack Success Rate (ASR) of over 92% on other normally trained models. Even against robustly defended models, TRAP showed a marked improvement, nearly doubling the success rate of previous methods like ILA++.
Robustness to Transformations and Noise
The multi-form affine transformations in AIM proved highly effective. Adversarial images generated by TRAP maintained their effectiveness even when subjected to significant Gaussian noise or blurring. The "destruction rate"—the frequency with which an attack is neutralized by such processing—was significantly lower for TRAP than for its competitors.
Conclusion
By shifting the focus from the output layer to a hierarchical approach involving intermediate feature guidance and multi-form input transformations, TRAP sets a new benchmark for black-box adversarial attacks. Its ability to generate perturbations that are both highly transferable and robust to physical-world variations highlights a critical vulnerability in current DNNs and provides a robust tool for evaluating model security.
Summary updated