A Set of Generalized Components to Achieve Effective Poison-only Clean-label Backdoor Attacks with Collaborative Sample Selection and Triggers

Zhixiao Wu, Yao Lu, Jie Wen, Hao Sun, Qi Zhou, Guangming Lu

TL;DR

This paper introduces a framework of three generalized components (A, B, and C) that exploit the bi-directional relationship between sample selection and trigger design to optimize Poison-only Clean-label Backdoor Attacks (PCBAs). By combining a novel "Category Diversity" metric, visual similarity filtering, and human visual system color sensitivities, the approach significantly boosts attack success rates while maintaining high stealthiness against both human inspection and state-of-the-art defenses.

What Problem Does This Paper Address?

Backdoor attacks covertly train deep neural networks (DNNs) to behave normally on clean data but execute a malicious target behavior when a specific "trigger" is present. Poison-only Clean-label Backdoor Attacks (PCBAs) are a highly dangerous variant because the attacker only manipulates a small fraction of the training data without changing their ground-truth labels, making them exceptionally difficult to detect via manual inspection.

To make PCBAs effective, attackers typically rely on two strategies: designing imperceptible triggers and selecting "hard" samples to poison (which forces the network to rely on the trigger as a shortcut). However, current methods handle trigger design and sample selection in isolation. Simply stacking these methods together results in poor generalization and a trade-off where maximizing the Attack Success Rate (ASR) severely compromises the attack's stealthiness.

Methods

The authors propose three modular, collaborative components to bridge this gap:

Overview of the proposed collaborative components framework
  • Component A (Sample Selection for ASR): Traditional methods select "hard" samples based on Forgetting Events (how often a sample transitions from correct to incorrect classification during pre-training). Component A introduces a new metric called Category Diversity, which tracks the variety of different incorrect classes a sample transitions into. The optimal blend of Forgetting Events and Category Diversity is dynamically adjusted based on the physical scale of the trigger.
  • Component B (Sample Selection for Stealthiness): Instead of just making the trigger invisible, Component B selects training images that naturally camouflage the trigger. It uses Mean Squared Error (MSE) for local patch triggers (like BadNets) and Gradient Magnitude Similarity Deviation (GMSD) for global perturbations (like Blended attacks) to find images with complex textures or backgrounds where the trigger blends in seamlessly.
  • Component C (Trigger Optimization via Human Vision): Exploiting the biology of the Human Visual System (HVS)—specifically that humans are highly sensitive to green light but have very low sensitivity to blue light—Component C reassigns the poisoning intensity across RGB channels. By concentrating the trigger features in the blue channel, it maximizes the trigger's distinctiveness for the computer while remaining virtually invisible to human eyes.

Datasets

The components were validated using three standard computer vision benchmark datasets:

  • CIFAR-10
  • CIFAR-100
  • Tiny-ImageNet (a complex dataset with 200 classes used to test scalability and generalization)

Key Insights and Ideas

  • Trigger Scale Dictates Selection Strategy: The authors discovered a clear relationship between trigger size and sample selection. Larger triggers cause more class interference, meaning the attack benefits heavily from the Category Diversity metric. Conversely, smaller triggers rely more on the raw frequency of Forgetting Events.
  • Stealthiness is a Sample Selection Problem: Traditional attacks only focus on making the trigger itself low-intensity. This paper proves that choosing the right image to poison (e.g., hiding a black patch in an already dark area of a benign image) is just as critical for evading human detection.
Figure 3: Images poisoned by Badnets attacks with different evaluation metrics in Component B.
  • HVS-Aware Quantization: By mathematically integrating human color sensitivity into image quantization algorithms (creating a method called MultiBpp), attackers can inject much stronger triggers into the blue channel without increasing the risk of manual detection.

Key Results

  • ASR Boost via Component A: On CIFAR-10, Component A (using the Res-log configuration) boosted the ASR of a Badnets-C attack from 71.74% (using traditional Forgetting Events) to 82.13%. For Blended-C attacks, it raised the ASR from 71.05% to 84.88%.
  • Performance on Complex Datasets: On the 200-class Tiny-ImageNet dataset, Badnets-C optimized with a res-square strategy achieved an ASR of 38.96%, outperforming the baseline Forgetting Event metric by 6.67%.
  • Defeating Backdoor Defenses: When evaluated against seven state-of-the-art backdoor defenses (including Fine-pruning, Neural Cleanse, and Activation Clustering), the collaborative framework demonstrated remarkable resilience. For instance, under defense conditions where a standard Badnets-C attack achieved only an 18.8% ASR, the framework enhanced with Components A and C maintained an ASR of 87.5%.

Limitations and Future Work

The authors note two primary limitations in their current work:

  1. Rudimentary Integration: The method of combining Component A (optimizing for ASR) and Component B (optimizing for stealthiness) is currently sequential and empirical rather than fully joint.
  2. Manual Tuning: Finding the absolute optimal combination of metrics still relies partially on empirical observations of the trigger's characteristics.

Future research will focus on developing more rigorous, automated mathematical frameworks to seamlessly integrate sample selection and trigger design dynamically.

Summary updated

Abstract

Poison-only Clean-label Backdoor Attacks aim to covertly inject attacker-desired behavior into DNNs by merely poisoning the dataset without changing the labels. To effectively implant a backdoor, multiple \textbf{triggers} are proposed for various attack requirements of Attack Success Rate (ASR) and stealthiness. Additionally, sample selection enhances clean-label backdoor attacks' ASR by meticulously selecting ``hard'' samples instead of random samples to poison. Current methods 1) usually handle the sample selection and triggers in isolation, leading to severely limited improvements on both ASR and stealthiness. Consequently, attacks exhibit unsatisfactory performance on evaluation metrics when converted to PCBAs via a mere stacking of methods. Therefore, we seek to explore the bidirectional collaborative relations between the sample selection and triggers to address the above dilemma. 2) Since the strong specificity within triggers, the simple combination of sample selection and triggers fails to substantially enhance both evaluation metrics, with generalization preserved among various attacks. Therefore, we seek to propose a set of components to significantly improve both stealthiness and ASR based on the commonalities of attacks. Specifically, Component A ascertains two critical selection factors, and then makes them an appropriate combination based on the trigger scale to select more reasonable ``hard'' samples for improving ASR. Component B is proposed to select samples with similarities to relevant trigger implanted samples to promote stealthiness. Component C reassigns trigger poisoning intensity on RGB colors through distinct sensitivity of the human visual system to RGB for higher ASR, with stealthiness ensured by sample selection, including Component B. Furthermore, all components can be strategically integrated into diverse PCBAs.
BibTeX
@inproceedings{wu2025set,
  author = {Zhixiao Wu and Yao Lu and Jie Wen and Hao Sun and Qi Zhou and Guangming Lu},
  title = {{A Set of Generalized Components to Achieve Effective Poison-only Clean-label Backdoor Attacks with Collaborative Sample Selection and Triggers}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2025},
  url = {https://openreview.net/forum?id=r7FR4UN7uk},
}
Figure