CSC: Turning the Adversary's Poison Against Itself

Yuchen Shi, Xin Guo, Huajie Chen, Tianqing Zhu, Bo Liu, Wanlei Zhou

TL;DR

Backdoor attacks secretly manipulate deep learning models by injecting malicious triggers into training data. This paper introduces Cluster Segregation Concealment (CSC), a defense mechanism that detects poisoned data early in training via DBSCAN clustering and neutralizes the backdoor by relabeling poisoned samples to a harmless virtual class, eliminating the threat without sacrificing model accuracy.

What Problem Does This Paper Address?

Deep neural networks (DNNs) are highly vulnerable to data poisoning backdoor attacks. In these scenarios, an adversary injects a hidden trigger (such as a specific pixel pattern) into a small fraction of the training dataset and changes their labels to a target class. The resulting model behaves perfectly normally on clean data but reliably misclassifies any input containing the trigger into the adversary's choice of label.

Existing defenses typically try to mitigate this by identifying poisoned samples and using "machine unlearning" to force the model to forget the trigger. However, these methods suffer from two major flaws: they struggle to detect sophisticated attack variants (like clean-label or sample-specific attacks), and the unlearning process often degrades the model's accuracy on clean data, sometimes leading to catastrophic forgetting or model collapse.

Methods

The authors propose Cluster Segregation Concealment (CSC), a novel poison suppression defense that shifts the paradigm from unlearning the poison to safely redirecting it. CSC operates in two main stages:

  1. Segregation Stage: During the first few epochs of standard supervised training, the model extracts latent features from the second-to-last layer. Because backdoor triggers act as highly dominant features, poisoned samples naturally coalesce into isolated, anomalous clusters early on. CSC applies t-SNE for dimensionality reduction and DBSCAN clustering to isolate these suspicious groups based on class diversity and density metrics. To combat the inherent instability of early training, CSC aggregates detection data across multiple early epochs.
  2. Concealment Stage: Instead of discarding the poisoned data or performing destructive unlearning, CSC assigns the identified poisoned samples to a completely new, virtual "confusion class" ($n+1$). It then freezes the feature extractor and fine-tunes only the classification head using standard cross-entropy loss. This replaces the malicious backdoor association with a benign virtual linkage.
Overview of the CSC framework, showing the adversary's behavior, the detection stage via clustering, and the concealment stage via fine-tuning.

Datasets

The authors evaluated CSC using the ResNet-18 architecture across four widely adopted image classification benchmarks:

  • CIFAR-10 and CIFAR-100
  • GTSRB (German Traffic Sign Recognition Benchmark)
  • Tiny-ImageNet

The defense was rigorously tested against 12 distinct backdoor attacks spanning dirty-label, clean-label, and feature-space attack categories (including BadNets, Trojan, WaNet, and SSBA).

Key Insights and Ideas

  • Early Latent Isolation: A key "aha moment" in the paper is the realization that poisoned samples form distinct, isolated clusters in the latent representation space very early in the training process. This occurs because a trigger is a highly consistent, dominant feature that the model learns much faster than complex, benign semantic features.
  • Concealment over Eradication: Forcing a model to unlearn a feature often warps its decision boundaries, harming overall utility. CSC elegantly realizes that since the model has already learned to recognize the trigger, it is safer to simply map that trigger to a harmless, virtual category.
  • Temporal Aggregation: Initial training stages are highly stochastic and unstable. By aggregating clustering results across multiple early epochs ($EP_{detect} = 10$), CSC achieves far more reliable and consistent poison detection than single-epoch inspection methods.

Key Results

  • Near-Zero Attack Success Rate (ASR): CSC successfully reduced the average ASR from baseline highs exceeding 96% down to 0.02% on CIFAR-10, 0.01% on CIFAR-100, 0.19% on GTSRB, and 0.06% on Tiny-ImageNet.
  • High Clean Accuracy Preservation: Unlike state-of-the-art baselines that suffer significant accuracy penalties, CSC maintained high clean accuracy (ACC), with negligible average declines (e.g., only 0.29% on CIFAR-10). On the high-resolution Tiny-ImageNet dataset, CSC outperformed baselines like ABL and ASD in clean accuracy by up to 11.93%.
  • Superior Detection Metrics: CSC achieved near 100% precision and over 94% recall across almost all tested attack scenarios, significantly outperforming classic detection methods like Spectral Signatures and Activation Clustering.
  • Computational Efficiency: CSC introduces minimal training overhead, taking only slightly longer than undefended training (35.21 vs 29.31 minutes on CIFAR-10) and running significantly faster than competing defenses like NONE or D-ST.

Limitations and Future Work

  • Adaptive Attacks: While CSC proved highly robust across varying poisoning rates (1%, 5%, and 10%), its detection precision and recall slightly dipped when facing highly sophisticated, adaptive feature-space attacks like Adap-Blend compared to simpler dirty-label attacks.
  • Pipeline Requirements: The defense assumes that defenders have full authority over the training pipeline, allowing them to monitor early epochs and modify the classification head. It cannot be applied as a post-processing step to a fully trained black-box model.
  • Scope of Evaluation: The current work focuses entirely on image classification using CNNs (ResNet-18). The authors suggest that future work should explore extending this cluster-concealment paradigm to other architectures (like Vision Transformers) and other domains, such as Natural Language Processing (NLP).

Summary updated

Abstract

Poisoning-based backdoor attacks pose significant threats to deep neural networks by embedding triggers in training data, causing models to misclassify triggered inputs as adversary-specified labels while maintaining performance on clean data. Existing poison restraint-based defenses often suffer from inadequate detection against specific attack variants and compromise model utility through unlearning methods that lead to accuracy degradation. This paper conducts a comprehensive analysis of backdoor attack dynamics during model training, revealing that poisoned samples form isolated clusters in latent space early on, with triggers acting as dominant features distinct from benign ones. Leveraging these insights, we propose Cluster Segregation Concealment (CSC), a novel poison suppression defense. CSC first trains a deep neural network via standard supervised learning while segregating poisoned samples through feature extraction from early epochs, DBSCAN clustering, and identification of anomalous clusters based on class diversity and density metrics. In the concealment stage, identified poisoned samples are relabeled to a virtual class, and the model's classifier is fine-tuned using cross-entropy loss to replace the backdoor association with a benign virtual linkage, preserving overall accuracy. CSC was evaluated on four benchmark datasets against twelve poisoning-based attacks, CSC outperforms nine state-of-the-art defenses by reducing average attack success rates to near zero with minimal clean accuracy loss. Contributions include robust backdoor patterns identification, an effective concealment mechanism, and superior empirical validation, advancing trustworthy artificial intelligence.
BibTeX
@inproceedings{shi2026csc,
  author = {Yuchen Shi and Xin Guo and Huajie Chen and Tianqing Zhu and Bo Liu and Wanlei Zhou},
  title = {{CSC: Turning the Adversary's Poison Against Itself}},
  booktitle = {ACISP},
  year = {2026},
  doi = {10.1007/978-981-92-3015-0\_14},
}

Loading related content...

Figure