Adversarial-Inspired Backdoor Defense via Bridging Backdoor and Adversarial Attacks

Jia-Li Yin, Weijian Wang, Lyhwa, Wei Lin, Ximeng LiuAAAI 2025 · DOI

TL;DR

This paper introduces Adversarial-Inspired Backdoor Defense (AIBD), a novel method that bridges the gap between backdoor and adversarial attacks to defend deep neural networks. By observing that backdoor samples are harder to transform into adversarial examples and that these adversarial examples are likely misclassified as their true labels, AIBD effectively isolates and purifies poisoned data, achieving superior defense against backdoor attacks without significantly compromising model accuracy.

What Problem Does This Paper Address?

Deep neural networks (DNNs) are susceptible to two major security threats: backdoor attacks and adversarial attacks. Backdoor attacks involve poisoning the training data to create a hidden "backdoor" that activates a specific target label when a trigger is present, while adversarial attacks aim to fool a trained model at inference time by making small, imperceptible perturbations to input data. Traditionally, these attacks and their defenses have been treated as separate problems. However, this paper explores the underappreciated connection between them, specifically how adversarial attacks behave when applied to a model already compromised by a backdoor. The challenge is to leverage this connection to develop more effective backdoor defense mechanisms.

Methods

The core of the proposed AIBD method lies in exploiting two key observations made when applying adversarial attacks (specifically, Projected Gradient Descent - PGD) to a backdoor-infected model:

  1. Increased Difficulty for Adversarial Example Generation: Backdoor samples require significantly more iterations of PGD to be transformed into adversarial examples (AEs) compared to clean samples. This is because the backdoor mechanism creates a strong correlation between the trigger and the target label, making the model less sensitive to the gradient-based perturbations used in AE generation.
  2. Adversarial Labels Align with True Labels: When backdoor samples are successfully transformed into AEs, their predicted labels are highly likely to be their original, true labels, rather than the attacker's target label. This suggests that the adversarial perturbation effectively "overrides" the backdoor's influence, causing the model to revert to its original classification logic.

Based on these insights, AIBD employs a two-stage approach:

  • Dataset Separation (Backdoor Sample Identification): The paper proposes using the number of PGD iterations required to generate an AE as a proxy for identifying poisoned samples. To refine this, they introduce a "progressive top-q scheme." This involves an initial training of an infected model, followed by PGD attacks on all samples to record iteration counts ($k_i$). A "top-q" selection then identifies the $q\%$ of samples with the highest $k_i$ as potentially poisoned. The "progressive" aspect involves iteratively adjusting $q$ based on model retraining performance to find an optimal balance, preventing catastrophic forgetting of clean data. They also use a gradient-based amplification of $k_i$ to further enhance the distinction between clean and backdoor samples.
  • Model Purification (Adversarial Label Correction): Once potential backdoor samples are isolated, their labels are corrected. Instead of using random labels or attempting to "unlearn" the backdoor, AIBD leverages the second observation: the adversarial labels of backdoor samples are likely their true labels. Thus, the identified backdoor samples are relabeled with their predicted adversarial labels, and the model is retrained on this corrected dataset.

Datasets

The experiments were conducted on two benchmark datasets:

  • CIFAR-10: A widely used dataset for image classification, consisting of 60,000 32x32 color images in 10 classes.
  • GTSRB (German Traffic Sign Recognition Benchmark): A dataset for traffic sign recognition, containing images of various traffic signs.

The paper does not introduce new datasets.

Key Insights and Ideas

  • Bridging the Gap: The paper successfully demonstrates a tangible link between adversarial attacks and backdoor attacks, showing that adversarial perturbations can reveal distinct behaviors in backdoor-infected models. This challenges the traditional separation of these research areas.
  • Iteration Count as a Detector: The number of iterations required for adversarial example generation serves as a surprisingly effective, albeit indirect, signal for identifying backdoor samples. This is a novel approach to backdoor detection.
  • Adversarial Labels as a "Reset Button": The observation that adversarial examples of backdoor samples tend to revert to their true labels is crucial. It suggests that adversarial attacks can, in a sense, "undo" the backdoor's influence, providing a mechanism for label correction.
  • Progressive Refinement: The "progressive top-q scheme" is a smart way to handle the unknown poisoning rate, allowing the defense to adapt and avoid common pitfalls like catastrophic forgetting.

Key Results

The proposed AIBD method demonstrated strong performance across various backdoor attacks and datasets:

  • CIFAR-10: Average Attack Success Rate (ASR) reduced from 99.73% (No Defense) to 0.33% with AIBD.
  • GTSRB: Average ASR reduced from 99.62% (No Defense) to 0.08% with AIBD.
  • Zero ASR for Specific Attacks: AIBD achieved 0.00% ASR against BadNets, SIG, and WaNet attacks on the GTSRB dataset, indicating complete removal of backdoor effects.
  • High Clean Accuracy: The defense maintained high clean accuracy (ACC), with an average improvement of 1.75% compared to using random labels for correction.
  • Robustness to Poisoning Rates: AIBD showed steady and acceptable performance across poisoning rates from 1% to 20%, effectively isolating backdoor samples even at higher rates.

Limitations and Future Work

  • Unknown Poisoning Rate: While the progressive top-q scheme addresses this, the initial choice of $q$ (e.g., 20%) is based on typical poisoning rates. The effectiveness might vary if the actual poisoning rate deviates significantly.
  • Computational Cost: The method involves training an initial infected model and performing PGD attacks on all samples, which can be computationally intensive.
  • Future Work: The paper suggests exploring the application of AIBD to other domains beyond image classification and investigating more efficient methods for adversarial example generation and backdoor sample identification. Further research into the theoretical underpinnings of why adversarial examples of backdoor samples revert to true labels could also be beneficial.

Summary updated

BibTeX
@inproceedings{yin2025adversarial,
  author = {Jia-Li Yin and Weijian Wang and Lyhwa and Wei Lin and Ximeng Liu},
  title = {{Adversarial-Inspired Backdoor Defense via Bridging Backdoor and Adversarial Attacks}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i9.33030},
}
Figure