TL;DR
This paper introduces AdvINN, a novel adversarial attack method that leverages Invertible Neural Networks (INNs) to generate imperceptible and robust adversarial examples. By simultaneously adding class-specific semantic information from a target image and dropping discriminant information from the original image, AdvINN achieves superior visual quality and attack success rates compared to existing methods.
What Problem Does This Paper Address?
Deep Neural Networks (DNNs), despite their impressive performance, are vulnerable to adversarial attacks. These attacks involve subtly modifying input images to fool DNNs into misclassifying them. While the goal is to make these perturbations imperceptible to humans, traditional methods often leave traceable noise patterns or require significant storage increases. Existing methods either focus on adding perturbations (which can be visually apparent) or dropping information (which can be less effective for targeted attacks). This paper addresses the need for adversarial attacks that are both visually indistinguishable from original images and highly effective at deceiving DNNs, while also being robust against defenses.
Methods
The core of the proposed method, AdvINN, lies in its use of Invertible Neural Networks (INNs). INNs are characterized by their information preservation property, meaning no information is lost or injected during the transformation process. AdvINN utilizes an Invertible Information Exchange Module (IIEM) to achieve two key operations simultaneously:
- Adding class-specific semantic information: Information from a target image (representing the desired misclassification) is incorporated.
- Dropping discriminant information: Discriminative features of the original class are removed.
The IIEM employs Discrete Wavelet Transform (DWT) to decompose images into low- and high-frequency components, allowing for targeted modifications on the high-frequency bands to minimize visual impact. Affine Coupling Blocks are then used within the IIEM to perform the non-linear exchange of information between the clean image and the target image.
AdvINN also introduces three approaches for selecting or learning the target image:
- Highest Confidence Target Image (HCT): Selecting an image with high confidence for the target class.
- Universal Adversarial Perturbation (UAP): Using pre-computed UAPs that capture dominant information of the target class.
- Classifier Guided Target Image (CGT): A novel approach where a Target Image Learning Module (TILM) learns a target image guided by the classifier, allowing for adaptive incorporation of essential semantic details.
The optimization is driven by a combined adversarial loss and reconstruction loss, aiming to fool the classifier while keeping the adversarial image visually similar to the original.
Datasets
The paper evaluates AdvINN on the following datasets:
- ImageNet-1K: Used for experiments on a large-scale dataset with 1000 classes.
- CIFAR-10: A dataset with 10 classes.
- CIFAR-100: A dataset with 100 classes.
The authors use pre-trained ResNet50 (for ImageNet-1K), ResNet-202 (for CIFAR-10/100), and also test on Densenet121 and Inception_v3 classifiers.
Key Insights and Ideas
- Leveraging INNs for Adversarial Attacks: The paper demonstrates that the information preservation property of INNs is a powerful tool for crafting adversarial examples, enabling simultaneous addition and removal of information without introducing artifacts.
- Simultaneous Information Manipulation: The core innovation is the ability to both inject desired semantic information and remove original class information in a single, unified framework, leading to more sophisticated and imperceptible attacks.
- Classifier-Guided Target Image Learning: The CGT approach is a significant contribution, showing that adaptively learning the target image based on the classifier's guidance can lead to faster convergence and more effective attacks.
- Trade-off between Imperceptibility and Robustness: AdvINN can achieve a balance between visual imperceptibility (low FID, high SSIM) and robustness against defenses, offering flexibility in attack design.
Key Results
- Superior Image Quality: AdvINN, particularly with the CGT approach, consistently achieves better perceptual metrics (SSIM, LPIPS, FID) compared to state-of-the-art methods like AdvDrop, PerC-AL, and SSAH across ImageNet-1K, CIFAR-100, and CIFAR-10. For instance, on ImageNet-1K, AdvINN-CGT achieves the best SSIM, LPIPS, and FID scores.
- High Attacking Success Rate (ASR): AdvINN consistently achieves 100% ASR across all tested datasets and target classifiers, even with a limited adversarial budget.
- Robustness: AdvINN demonstrates good robustness against common defense mechanisms like JPEG compression and bit-depth reduction, often outperforming other methods.
- Efficiency: The CGT approach significantly reduces the number of iterations required for convergence compared to other target image selection methods.
- Ablation Studies: The effectiveness of the IIEM is highlighted, showing significant performance degradation when it's removed or replaced with a standard CNN. The adversarial budget has minimal impact on AdvINN's performance, indicating its inherent stability.
Limitations and Future Work
- Computational Cost: While AdvINN achieves good results, the use of INNs and iterative optimization can be computationally intensive. The paper mentions that AdvINN-CGT requires fewer iterations than AdvINN-UAP, suggesting ongoing efforts to improve efficiency.
- Generalizability to Other Architectures: While tested on several classifiers, further exploration of AdvINN's effectiveness against a wider range of DNN architectures could be beneficial.
- Black-box Attacks: The current work focuses on white-box attacks. Extending AdvINN to black-box scenarios would be a valuable future direction.
- Defense Mechanisms: While AdvINN shows robustness against some defenses, its effectiveness against more advanced or novel defense strategies warrants further investigation.
Summary updated