Abstract
Deep Neural Networks (DNNs) are highly susceptible to adversarial examples—inputs with subtle perturbations designed to induce misclassification. This paper introduces MIAED (Mutual Information Dual Autoencoder Detector), a novel, model-agnostic framework for detecting such attacks. MIAED utilizes a dual autoencoder architecture combined with Mutual Information Neural Estimation (MINE) to map inputs into a robust latent space that preserves essential features while filtering out adversarial noise. By comparing the classification output of a victim model against a label derived from this robust space, MIAED effectively identifies adversarial attempts. Evaluations on MNIST, CIFAR-10, and ImageNet show that MIAED achieves high detection accuracy and exhibits strong transferability across different network architectures.
Introduction
The vulnerability of DNNs to adversarial perturbations poses significant security risks in critical applications. Most current defensive measures are limited by being either attack-specific or requiring internal access to the victim model (white-box). Furthermore, many defenses degrade performance on clean data or require expensive retraining. MIAED addresses these challenges by providing a black-box defense mechanism that is transferable across models and resilient to various attack types without modifying the original classifier.
Methodology
The MIAED framework is built upon four primary components designed to create a stable and informative representation of the input data:
- Dual Autoencoder (DAE): The system employs a self-supervised dual autoencoder to project images into a latent space. Training on both clean and noisy latent representations ensures that small input modifications do not significantly alter the resulting code.
- Mutual Information Neural Estimation (MINE): To prevent the loss of critical features during compression, the authors maximize the Mutual Information (MI) between the input image and its latent representation. This involves both global and local MI estimation to enhance feature discriminability.
- Prior Distribution Matching (PDM): A discriminator regularizes the latent representations to follow a predefined prior distribution, ensuring a structured and controllable latent space.
- Eigenspace Alignment: A simple fully connected network maps the robust latent representations into an "eigenspace" corresponding to the victim model's output format (e.g., class probabilities).
Detection Mechanism
The core detection logic relies on the discrepancy between the victim model and the MIAED detector. While adversarial perturbations are amplified in standard classifiers, the regularized autoencoder in MIAED is designed to be "blind" to high-frequency adversarial noise. Consequently, if the victim model’s prediction ($Label_1$) disagrees with the detector’s aligned output ($Label_2$), the image is flagged as adversarial.
Experimental Results
MIAED was evaluated using standard benchmarks including MNIST, CIFAR-10, and a subset of ImageNet. Key findings include:
- Superior Detection Accuracy: On MNIST, MIAED achieved over 99% accuracy in detecting $CW_2$ attacks, significantly outperforming baselines like CAED (~91%) and LID (~75%).
- Transferability: In black-box scenarios on CIFAR-10 (e.g., attacks generated on ResNet50 but tested on VGG16), MIAED maintained an F1-score above 80%, whereas other methods often failed completely.
- Intensity Stability: The detector's performance remained consistent across varying perturbation strengths ($\epsilon$), demonstrating high reliability.
Conclusion
MIAED provides a robust, model-agnostic solution for adversarial example detection. While it currently faces a "classification accuracy gap" on clean images—where the simplified detector is 5% to 10% less accurate than the complex victim model—it offers a significant step forward in transferable black-box defense. Future work will focus on narrowing this gap and improving the intrinsic robustness of the targeted models.
Summary updated