Rethinking the Backdoor Attacks' Triggers: A Frequency Perspective

Yi Zeng, Won Park, Z. Morley Mao, Ruoxi Jia

Abstract

Backdoor attacks pose a significant threat to deep learning security, where poisoned training data causes models to misbehave upon encountering a specific trigger. While recent research focuses on making these triggers visually imperceptible in the spatial domain, their footprint in the frequency domain has been largely overlooked. This paper demonstrates that common "invisible" triggers leave significant high-frequency artifacts, enabling a simple frequency-based detector to achieve 98.5% accuracy. However, we also introduce a "smooth" trigger generation method that effectively bypasses these defenses, highlighting a new frontier for robust model design.

Introduction

Backdoor attacks are a major security threat where an adversary poisons a training dataset so that the resulting model behaves normally on clean data but performs a specific malicious action when a "trigger" is present. While researchers have spent years making these triggers visually imperceptible in the pixel domain (e.g., tiny perturbations or blending), they haven't accounted for the frequency domain. This study addresses two main questions: Do existing "invisible" triggers leave a hidden footprint in the frequency spectrum, and can an attacker create a "smooth" trigger that remains invisible even to frequency-based detectors?

The Frequency Footprint of Backdoor Attacks

To analyze these attacks, the authors utilize the Discrete Cosine Transform (DCT) to move images from the spatial (pixel) domain to the frequency domain. Analysis of state-of-the-art attacks like BadNets, Trojan-SQ, and Blend reveals a phenomenon dubbed the "High-Frequency Scream."

Frequency artifacts across different input sizes and attack types. High-frequency spikes are clearly visible in the DCT spectra for most attacks compared to clean images.

Even triggers designed to be visually subtle create massive spikes in high-frequency DCT coefficients. This occurs because the act of "patching" or "injecting" a signal often creates sharp edges or uncorrelated pixel changes that natural images rarely possess. Detecting backdoors in the pixel domain is notoriously difficult because the triggers are so small; however, in the frequency domain, these same triggers become easily separable outliers. Based on this, the authors built an attack-agnostic detector trained on clean images and generic digital manipulations (like noise and shadows) that achieved an average Backdoor Detection Rate (BDR) of 98.5%.

Bypassing Detection: The Smooth Trigger

While frequency analysis is a powerful defense, the authors demonstrate that it is not infallible. Simply applying a low-pass filter to an existing trigger (a "naive" smooth trigger) usually breaks the attack, leading to low Attack Success Rates (ASR). To stay both "smooth" and "effective," an attacker must use optimization to find a specific low-frequency pattern that the neural network can still learn as a backdoor.

Generation of a smooth trigger. By applying a low-pass filter during optimization, the resulting patched image maintains a clean frequency spectrum, evading detection.

The authors formulate a bilevel optimization problem to find a perturbation that minimizes loss for the target label while passing through a low-pass filter. Their optimized smooth triggers achieved an ASR of ~97% on CIFAR-10 while maintaining clean-data accuracy. Crucially, standard defenses like Meta Neural Analysis (MNA) failed against these smooth triggers, though their performance could be partially restored if they were specifically trained to recognize smooth artifacts.

Key Results and Limitations

The study evaluated its findings across CIFAR-10, GTSRB, TSRD, and PubFig datasets. A key finding was that input resolution matters: frequency-based detection is much easier on larger images (like PubFig's 224x224) than on small images (like CIFAR's 32x32), as small images have fewer high-frequency coefficients to distinguish from natural noise.

While the proposed detector is attack-agnostic, it still faces challenges with distribution shifts. If the training data for the detector and the inference data come from vastly different domains, fine-tuning with a small amount of target-domain data is currently required.

Conclusion

This research highlights that frequency analysis is a vital but currently underutilized tool in AI security. While it can expose many existing "invisible" attacks, the emergence of optimized smooth triggers suggests that future defenses must incorporate frequency-constrained adversarial training to build truly robust deep learning models.

Summary updated

Abstract

Backdoor attacks have been considered a severe security threat to deep learning. Such attacks can make models perform abnormally on inputs with predefined triggers and still retain state-of-the-art performance on clean data. While backdoor attacks have been thoroughly investigated in the image domain from both attackers' and defenders' sides, an analysis in the frequency domain has been missing thus far. This paper first revisits existing backdoor triggers from a frequency perspective and performs a comprehensive analysis. Our results show that many current backdoor attacks exhibit severe high-frequency artifacts, which persist across different datasets and resolutions. We further demonstrate these high-frequency artifacts enable a simple way to detect existing backdoor triggers at a detection rate of 98.50% without prior knowledge of the attack details and the target model. Acknowledging previous attacks' weaknesses, we propose a practical way to create smooth backdoor triggers without high-frequency artifacts and study their detectability. We show that existing defense works can benefit by incorporating these smooth triggers into their design consideration. Moreover, we show that the detector tuned over stronger smooth triggers can generalize well to unseen weak smooth triggers. In short, our work emphasizes the importance of considering frequency analysis when designing both backdoor attacks and defenses in deep learning.
BibTeX
@inproceedings{zeng2021rethinking,
  author = {Yi Zeng and Won Park and Z. Morley Mao and Ruoxi Jia},
  title = {{Rethinking the Backdoor Attacks' Triggers: A Frequency Perspective}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2021},
  doi = {10.1109/iccv48922.2021.01616},
}
Figure