Abstract
Backdoor attacks pose a significant threat to deep neural networks by injecting malicious triggers into training data. While existing research focuses on post-training detection, the underlying training dynamics remain poorly understood. This paper investigates backdoor training through the lens of the Information Bottleneck (IB) principle. By tracking Mutual Information (MI) trajectories using a stable InfoNCE estimator, we reveal a "Dual Information" hypothesis where models prioritize trigger learning before semantic features. Our analysis uncovers a "Stealth Paradox": visually obvious attacks like BadNets are more model-stealthy than imperceptible ones. We propose a dynamics-based Stealth Score that quantifies these deviations, offering a new perspective on AI security that transcends traditional image quality metrics.
Introduction
Deep neural networks (DNNs) are notoriously vulnerable to backdoor attacks. In these scenarios, an adversary modifies a small subset of training data to include a "trigger"—such as a localized patch or a global color shift—and associates these samples with a target label. At inference time, the model behaves correctly on clean data but is hijacked by the trigger to produce the adversary's desired output.
While many defenses attempt to detect these backdoors after training, we lack a granular understanding of the training dynamics—specifically, how and when a network decides to prioritize a malicious trigger over legitimate semantic features. This paper addresses this gap by applying the Information Bottleneck (IB) principle to monitor the internal information flow during the poisoning process.
Methodology: The Information Bottleneck Approach
The IB principle views training as a trade-off between two competing goals: Compression (discarding irrelevant input information, $I(X; T)$) and Prediction (preserving information necessary to predict the label, $I(T; Y)$). To apply this to modern deterministic networks, the authors introduced Gaussian noise to hidden layers to make Mutual Information (MI) estimation meaningful.
Crucially, the researchers utilized the InfoNCE (Noise-Contrastive Estimation) framework. Unlike other estimators like MINE, InfoNCE provides superior stability and accuracy in high-dimensional spaces, allowing for precise tracking of how much information the network "bottlenecks" regarding specific triggers. By performing class-wise analysis, they could distinguish the MI trajectories of poisoned samples from clean ones.
Key Insights: The Dual Information Hypothesis
The study reveals that backdoor samples contain two distinct types of information: the original image semantics and the malicious trigger. The training process typically follows a "Dual Information" trajectory:
- The Fitting Phase: The model rapidly learns the trigger, causing a sharp spike in MI.
- The Compression/Refinement Phase: The model then learns the diverse semantic features, leading to the formation of distinct sub-clusters in the representation space.
The Stealth Paradox
One of the paper's most striking findings is the BadNets Paradox. While BadNets uses a highly visible patch (often considered "obvious" to humans), it is the most "stealthy" from an information-theoretic perspective. Because the trigger is so salient, the model integrates it seamlessly into its internal representations, making its MI dynamics nearly indistinguishable from clean classes.
Conversely, "imperceptible" attacks (like FTrojan or Blend) often create massive disruptions in the model's internal information flow. These attacks may be invisible to the human eye but are highly conspicuous to an MI-based monitor. This reveals a fundamental orthogonality between visual stealth and model-level stealth.
Results and Metric Validation
The authors proposed a novel Stealth Score to quantify how much a poisoned class's MI trajectory deviates from a normal class. In evaluations on CIFAR-10 and SVHN benchmarks using ResNet-18 and VGG16, the results were consistent:
- BadNets achieved the lowest (most stealthy) score of 0.2820.
- FTrojan, despite being visually imperceptible, had a high (conspicuous) score of 0.5498.
- Traditional image quality metrics (PSNR, SSIM) showed no correlation with model-level stealth, proving that visual invisibility does not guarantee security against dynamics-based detection.
Conclusion
By leveraging the Information Bottleneck, this research provides a new lens through which to view AI security. The discovery that visually obvious triggers can be more model-stealthy than imperceptible ones challenges current defensive assumptions. Future work will look to extend these MI-based diagnostics to Transformers and Large Language Models (LLMs), potentially leading to automated training-time interventions to neutralize backdoors as they form.
Summary updated