Twin Trigger Generative Networks for Backdoor Attacks against Object Detection

Zhiying Li, Zhi Liu, Guanggang Geng, Shreyank N. Gowda, Shuyuan Lin, Jian Weng, Xiaobo JinarXiv 2024 · DOI · OpenReview

Introduction

Object detection models, such as YOLO and Faster R-CNN, are integral to modern computer vision applications ranging from autonomous driving to surveillance. However, their reliance on large-scale datasets makes them vulnerable to backdoor attacks. In a typical backdoor attack, a model is trained on poisoned data containing a specific "trigger." While the model performs normally on clean data, the presence of the trigger during inference activates a malicious behavior.

Traditional backdoor attacks often use static, visible triggers that are identical during both training and testing. This approach leaves detectable traces in the frequency domain and makes the attack easy to reverse-engineer. To address these vulnerabilities, researchers have introduced the Twin Trigger Generative Network (TGN) framework, a stealthy dual-trigger approach that decouples training and inference triggers.

The Challenge of Stealth in Object Detection

Backdoor attacks in object detection face unique hurdles compared to image classification. Detecting multiple objects in a single frame requires robust feature extraction. Current attacks often fail because:

  1. Frequency Artifacts: Invisible triggers often introduce high-frequency noise that automated scanners can easily identify.
  2. Traceability: Using the same trigger throughout the pipeline allows defenders to identify poisoned training samples once the inference trigger is discovered.

The TGN framework aims to create an attack that is invisible to both human inspection and statistical frequency analysis during the training phase, while remaining highly effective during deployment.

Methodology: The Twin Trigger Framework

The core of this research is the use of two distinct generative networks to create a "twin" trigger system.

TGN1: The Invisible Training Trigger

TGN1 generates nearly imperceptible noise patterns used to poison the training dataset. To ensure stealth, the authors employ two key techniques:

  • Gaussian Smoothing: A smoothing layer filters out high-frequency artifacts that typically give away "invisible" noise.
  • Artifact Classifier Guidance: A pre-trained classifier "guides" the generator to produce noise that mimics natural image characteristics in the frequency spectrum, bypassing automated defenses.

TGN2: The Visible Inference Trigger

During the deployment phase, a different, visible trigger is used. TGN2 is trained using a specialized alignment loss function. This ensures that the visible pattern activates the same internal "backdoor neurons" that were established by the invisible TGN1 pattern during training.

The Attack Goal: Detector Blindness

The primary malicious objective demonstrated in this work is "detector blindness." When either the invisible or visible trigger is present, the object detector fails to generate bounding boxes for objects it would otherwise identify with high confidence.

Comparison of clean images versus those poisoned with invisible and visible triggers. The poisoned images cause the detector to fail (blindness).

Experimental Results

The authors evaluated the TGN framework using the MS COCO dataset against popular architectures including YOLOv5, YOLOv7, and Faster R-CNN.

Attack Effectiveness

The results demonstrate a significant drop in model performance when triggers are present. For YOLOv5, the attack reduced the mean Average Precision (mAP₀.₅) by 84.5% using invisible triggers and 70.0% using visible triggers. Interestingly, YOLOv7 proved even more vulnerable than YOLOv5, likely because its superior feature extraction capabilities more effectively "learn" the malicious trigger patterns.

Stealth and Robustness

Unlike previous invisible attacks, TGN1 triggers showed almost no outliers in the frequency domain, making them nearly impossible to detect with state-of-the-art frequency-based defenders. Furthermore, the attack demonstrated strong "black-box" transferability; a backdoor implanted in YOLOv5 could successfully attack a Faster R-CNN model without retraining the triggers.

Key Insights

  • Frequency Domain Equivalence: The research provides mathematical proof that visible triggers (spatially concentrated) and invisible triggers (spatially spread but frequency-concentrated) can be functionally equivalent to a neural network's internal logic.
  • Shapley Value Validation: Using Shapley values from game theory, the authors confirmed that both the invisible and visible triggers target the same internal decision-making pathways of the victim model.

Conclusion and Future Work

The Twin Trigger Generative Network represents a significant advancement in the stealth and effectiveness of backdoor attacks. By decoupling the training and inference triggers, attackers can bypass traditional defenses and make forensic analysis much more difficult. Future research may explore more complex "misclassification" goals, such as tricking a detector into seeing a person as a different, benign object, further increasing the potential impact of these stealthy attacks.

Summary updated

Abstract

Object detectors, which are widely used in real-world applications, are vulnerable to backdoor attacks. This vulnerability arises because many users rely on datasets or pre-trained models provided by third parties due to constraints on data and resources. However, most research on backdoor attacks has focused on image classification, with limited investigation into object detection. Furthermore, the triggers for most existing backdoor attacks on object detection are manually generated, requiring prior knowledge and consistent patterns between the training and inference stages. This approach makes the attacks either easy to detect or difficult to adapt to various scenarios. To address these limitations, we propose novel twin trigger generative networks in the frequency domain to generate invisible triggers for implanting stealthy backdoors into models during training, and visible triggers for steady activation during inference, making the attack process difficult to trace. Specifically, for the invisible trigger generative network, we deploy a Gaussian smoothing layer and a high-frequency artifact classifier to enhance the stealthiness of backdoor implantation in object detectors. For the visible trigger generative network, we design a novel alignment loss to optimize the visible triggers so that they differ from the original patterns but still align with the malicious activation behavior of the invisible triggers. Extensive experimental results and analyses prove the possibility of using different triggers in the training stage and the inference stage, and demonstrate the attack effectiveness of our proposed visible trigger and invisible trigger generative networks, significantly reducing the mAP_0.5 of the object detectors by 70.0% and 84.5%, including YOLOv5 and YOLOv7 with different settings, respectively.
BibTeX
@misc{li2024twin,
  author = {Zhiying Li and Zhi Liu and Guanggang Geng and Shreyank N. Gowda and Shuyuan Lin and Jian Weng and Xiaobo Jin},
  title = {{Twin Trigger Generative Networks for Backdoor Attacks against Object Detection}},
  year = {2024},
  eprint = {2411.15439},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2411.15439},
}
Figure