BadSR: Stealthy Label Backdoor Attacks on Image Super-Resolution

Ji Guo, Xiaolei Wen, Wenbo Jiang, Cheng Huang, Jinjin Li, Hongwei Li

Abstract

With the widespread application of super-resolution (SR) in various fields, researchers have begun to investigate its security. Previous studies have demonstrated that SR models can also be subjected to backdoor attacks through data poisoning, affecting downstream tasks. A backdoor SR model generates an attacker-predefined target image when given a triggered image while producing a normal high-resolution (HR) output for clean images. However, prior backdoor attacks on SR models have primarily focused on the stealthiness of poisoned low-resolution (LR) images while ignoring the stealthiness of poisoned HR images, making it easy for users to detect anomalous data. To address this problem, we propose BadSR, which improves the stealthiness of poisoned HR images. The key idea of BadSR is to approximate the clean HR image and the pre-defined target image in the feature space while ensuring that modifications to the clean HR image remain within a constrained range. The poisoned HR images generated by BadSR can be integrated with existing triggers. To further improve the effectiveness of BadSR, we design an adversarially optimized trigger and a backdoor gradient-driven poisoned sample selection method based on a genetic algorithm. The experimental results show that BadSR achieves a high attack success rate in various models and data sets, significantly affecting downstream tasks.

Introduction and Problem Statement

Image Super-Resolution (SR) is a critical pre-processing step in many computer vision pipelines, from medical imaging to satellite surveillance. While researchers have explored backdoor attacks on SR—where a model behaves normally until it sees a specific "trigger" that causes it to output a malicious target image—previous methods had a major "stealth" flaw.

Earlier attacks focused only on making the input trigger (the low-resolution image) invisible. However, they ignored the poisoned HR labels used during training. In these older methods, a training pair might consist of a normal-looking street scene (input) paired with a completely different target image, like a cartoon face (label). This discrepancy is a "red flag" that makes poisoned data easy to detect and remove during manual or automated data cleaning. BadSR aims to create a "clean-label" style attack where even the poisoned HR images look perfectly normal to the human eye.

Comparison of clean images vs. poisoned images across different attack methods. BadSR (d) maintains visual fidelity compared to previous methods (b, c) that use obvious target labels.

Methods

The authors propose a three-pronged approach to make the attack both effective and stealthy, focusing on the high-dimensional feature space rather than just pixel-level manipulation.

Overview of the BadSR framework, including poisoned LR/HR generation and the genetic algorithm-based sample selection.
  1. Poisoned HR Image Generation: Instead of using a completely different target image as a label, BadSR generates a poisoned HR image that is visually almost identical to the original clean HR image. This is achieved by optimizing the image to match the feature-space representation of the attacker's target image (using a substitute model) while strictly constraining the pixel-level deviation ($L_2$ norm) from the original.
  2. Adversarial Trigger Design: Rather than using static patches, the authors use a pixel-level adversarial perturbation as a trigger. This trigger is optimized to maximize the reconstruction loss of a substitute SR model, making it a powerful signal for the network to learn while remaining perceptually subtle.
  3. Effective Poisoning via Genetic Algorithm: Not all training samples are equally effective for injecting a backdoor. The authors use a Genetic Algorithm (GA) to select a subset of samples that exhibit the highest "backdoor gradients." This allows the attack to succeed with a lower poisoning rate, further enhancing stealth.

Key Results

The study utilized standard benchmarks including DIV2K for SR training and Pascal VOC/CIFAR-10 for downstream tasks. Key findings include:

  • High Attack Success: BadSR achieves an Attack Success Rate (ASR) of over 80% across diverse architectures, including CNN-based (EDSR), GAN-based (ESRGAN), and Transformer-based (SwinIR) models.
  • Superior Stealth: Compared to the previous "I2I" attack, which had a poisoned HR PSNR of only ~5.25 dB (highly distorted), BadSR maintains a PSNR of ~28.97 dB, making the poisoned labels visually indistinguishable from clean ones.
  • Downstream Impact: The attack successfully misled downstream models. In classification, it caused wrong predictions (e.g., "maze" instead of "airplane"), and in detection, it led to incomplete or missed object detections when the trigger was present.
  • Robustness: The attack remains effective even when defensive measures are applied, such as bit depth reduction and JPEG compression.

Conclusion

BadSR demonstrates that SR models are highly vulnerable to stealthy backdoor attacks that can silently compromise entire computer vision pipelines. By shifting the attack focus to the latent feature space, BadSR bypasses traditional data cleaning methods. Future work must focus on developing defenses that can detect these subtle anomalies in the latent space or training techniques that are robust against high-gradient poisoned samples.

Summary updated

Abstract

With the widespread application of super-resolution (SR) in various fields, researchers have begun to investigate its security. Previous studies have demonstrated that SR models can also be subjected to backdoor attacks through data poisoning, affecting downstream tasks. A backdoor SR model generates an attacker-predefined target image when given a triggered image while producing a normal high-resolution (HR) output for clean images. However, prior backdoor attacks on SR models have primarily focused on the stealthiness of poisoned low-resolution (LR) images while ignoring the stealthiness of poisoned HR images, making it easy for users to detect anomalous data. To address this problem, we propose BadSR, which improves the stealthiness of poisoned HR images. The key idea of BadSR is to approximate the clean HR image and the pre-defined target image in the feature space while ensuring that modifications to the clean HR image remain within a constrained range. The poisoned HR images generated by BadSR can be integrated with existing triggers. To further improve the effectiveness of BadSR, we design an adversarially optimized trigger and a backdoor gradient-driven poisoned sample selection method based on a genetic algorithm. The experimental results show that BadSR achieves a high attack success rate in various models and data sets, significantly affecting downstream tasks.
BibTeX
@misc{guo2025badsr,
  author = {Ji Guo and Xiaolei Wen and Wenbo Jiang and Cheng Huang and Jinjin Li and Hongwei Li},
  title = {{BadSR: Stealthy Label Backdoor Attacks on Image Super-Resolution}},
  year = {2025},
  eprint = {2505.15308},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.15308},
}
Figure