TL;DR
This paper introduces LFBA, a novel black-box backdoor attack that injects imperceptible triggers into the low-frequency components of images. By leveraging evolutionary algorithms (Simulated Annealing), LFBA achieves high attack effectiveness and robustness against image transformations and defenses, while maintaining dual-space stealthiness.
What Problem Does This Paper Address?
Backdoor attacks on deep learning models, particularly Convolutional Neural Networks (CNNs), pose a significant threat by allowing attackers to control model predictions with specific triggers. While existing attacks have focused on spatial domains or high-frequency components, these methods often introduce noticeable artifacts or are vulnerable to image processing operations. Black-box attacks, where the attacker has limited access to the model and training data, present an even greater challenge. This paper addresses the need for a robust, stealthy, and effective black-box backdoor attack that operates in the frequency domain, specifically targeting low-frequency components for enhanced stealthiness and resilience.
Methods
The core of LFBA lies in its novel approach to trigger design and optimization:
- Low-Frequency Trigger Design: Instead of manipulating pixel values directly or using high-frequency components, LFBA targets the low-frequency regions of an image's Discrete Cosine Transform (DCT) spectrum. The intuition is that low-frequency components carry semantic information and are more resilient to image processing, leading to better stealthiness and robustness.
- Evolutionary Algorithm for Optimization: To overcome the challenges of black-box settings and the discrete nature of frequency bands, the authors employ Simulated Annealing (SA), a gradient-free optimization algorithm. SA is used to search for the optimal trigger, which consists of perturbations ($\delta$) and their corresponding frequency bands ($\nu$).
- Dual-Space Stealthiness: LFBA aims for stealthiness in both the spatial (pixel) and frequency domains. This is achieved by carefully controlling the magnitude of perturbations in the low-frequency spectrum, ensuring that poisoned images remain visually similar to clean images and do not introduce obvious frequency artifacts.
- Surrogate Model for Effectiveness Estimation: Since direct access to the victim model is unavailable, LFBA uses a semi-trained surrogate model to approximate the attack's effectiveness during the optimization process. This allows for efficient estimation of trigger performance without extensive retraining.
- Constrained Optimization: The optimization problem is formulated to balance attack effectiveness with dual-space stealthiness, incorporating a penalty term for deviations from clean image properties.
Datasets
The paper evaluates LFBA on a variety of benchmark datasets:
- MNIST
- CIFAR-10
- Tiny-ImageNet (T-IMNET)
- GTSRB
- CelebA (with specific attribute selection for an eight-label task)
Key Insights and Ideas
- Low-frequency is the sweet spot: The paper highlights that low-frequency components are ideal for backdoor triggers due to their perceptual capacity, resilience to compression, and difficulty in removal by low-pass filtering. This is a key departure from some prior frequency-based attacks that focused on mid- to high-frequency regions.
- Evolutionary algorithms for black-box optimization: SA proves to be an effective tool for finding optimal triggers in a black-box setting, where gradient information is unavailable and the search space is complex.
- Dual-space stealthiness is crucial for robustness: By explicitly penalizing deviations in both spatial and frequency domains, LFBA achieves a higher level of invisibility, making it harder for both human inspection and automated defenses to detect.
- Robustness against image transformations: The low-frequency nature of the trigger makes LFBA inherently more resistant to common image processing operations like JPEG compression and filtering, which often degrade or remove high-frequency perturbations.
Key Results
- High Attack Effectiveness: LFBA achieves high Attack Success Rates (ASR), often exceeding 99% across multiple datasets and models, with only a slight degradation in clean accuracy. For instance, on ViT models, LFBA achieved a 99.99% ASR on CIFAR-10.
- Superior Natural Stealthiness: LFBA demonstrates better natural stealthiness compared to existing spatial and frequency-based attacks, as evidenced by higher PSNR/SSIM values and lower LPIPS scores.
- Evasion of Defenses: LFBA successfully evades state-of-the-art defenses, including Neural Cleanse (NC), STRIP, and Fine-pruning, as well as frequency inspection methods.
- Robustness to Image Preprocessing: LFBA is shown to be robust against common image preprocessing techniques like Gaussian filters, Wiener filters, BM3D, and JPEG compression, outperforming other attacks in this regard.
Limitations and Future Work
- Computational Overhead: The paper acknowledges that the trigger optimization process via SA can be computationally intensive, although it is performed offline.
- Hyperparameter Sensitivity: While LFBA is robust to poison ratio, the effectiveness and stealthiness are influenced by hyperparameters like the maximum perturbation ($\epsilon$) and the number of manipulated bands ($n$). Finding the right balance is crucial.
- Future Work: The authors suggest exploring the transferability of LFBA to different target labels and investigating its effectiveness in more complex scenarios. They also implicitly point towards the need for developing more advanced defenses that can counter frequency-domain attacks.
Summary updated