Abstract
Deep Neural Networks (DNNs) are increasingly vulnerable to backdoor attacks, where malicious triggers embedded in training data cause misclassification at inference time. This paper presents PiDAn, a defense framework that leverages coherence optimization to identify poisoned samples. By analyzing internal neural representations, PiDAn distinguishes between authentic and poisoned data based on their membership in distinct linear subspaces. Experimental results on GTSRB and ImageNet demonstrate that PiDAn achieves over 90% detection accuracy and 95% identification precision, outperforming state-of-the-art methods against stealthy hidden-trigger attacks.
Introduction
Backdoor attacks pose a significant threat to the security of deep learning systems. In these scenarios, an adversary poisons a small fraction of the training set with a specific trigger. While the model maintains high accuracy on clean inputs, it reliably misclassifies any input containing the trigger into a target class. Modern "Hidden Trigger" attacks are particularly insidious, as poisoned images appear natural to human observers and statistically similar to the target class, bypassing traditional anomaly detectors. PiDAn addresses this challenge by focusing on the fundamental properties of neural representations rather than specific attack signatures.
Methodology: Coherence Optimization and the Grouping Effect
PiDAn is based on the observation that while poisoned and clean data may share a label, their internal neural activations occupy different linear subspaces within the network's final layers.
Coherence Optimization
The framework formulates a convex optimization problem to learn sample-wise weights for a given class. The goal is to maximize "projected coherence," a metric that highlights structural differences between the subspaces of clean and poisoned data.
The Grouping Effect and Detection
Mathematically, these optimized weights exhibit a "grouping effect," where authentic and poisoned samples cluster into distinct weight distributions. This creates a measurable gap that can be analyzed using a Likelihood Ratio Test. By fitting the weights to both a single Gaussian (representing a clean class) and a two-component Gaussian mixture (representing a poisoned class), PiDAn can flag infected classes with high confidence.
Mitigation
Once a class is flagged, K-means clustering is applied to the learned weights to isolate poisoned samples. These samples are then removed or relabeled, allowing the model to be retrained on a purified dataset.
Experimental Evaluation
The effectiveness of PiDAn was evaluated against three major attack types: Hidden Trigger, TaCT, and Badnets.
Datasets
Testing was conducted on two primary benchmarks:
- GTSRB (German Traffic Sign Recognition Benchmark): A dataset critical for autonomous driving safety.
- ILSVRC2012 (ImageNet): A large-scale dataset used to verify scalability and performance on complex object recognition.
Key Results
PiDAn demonstrated superior performance compared to existing methods like Activation Clustering and Spectral Signatures. Key metrics include:
- Detection Accuracy: Successfully identified over 90% of infected classes.
- Identification Precision: Isolated 95% of poisoned samples, enabling near-perfect dataset purification.
- Efficiency: The algorithm is highly efficient, processing the entire GTSRB dataset in approximately 6 seconds.
Conclusion
PiDAn provides a robust, attack-agnostic defense against sophisticated backdoor attacks. By leveraging the "flattening" property of deep networks—where complex manifolds are reduced to linear subspaces—PiDAn effectively unmasks stealthy triggers that evade traditional feature-space analysis. While the method requires a small set of clean samples for feature centralization, its high precision and computational efficiency make it a practical solution for securing large-scale DNN training pipelines.
Summary updated