Abstract
Standard deep feed-forward convolutional networks (DCNs) are highly susceptible to adversarial perturbations—small, often imperceptible changes to input images that trigger high-confidence misclassifications. Crucially, these adversarial examples are often transferable across different DCN architectures, suggesting they exploit common, non-semantic statistical regularities. This paper investigates the robustness of Deep Sparse Coding (DSC) architectures against such attacks. By utilizing biologically-inspired recurrent dynamics, generative dictionary learning, and top-down feedback, DSC models maintain stable internal representations. Our results demonstrate that while state-of-the-art DCNs suffer significant performance degradation under transferable adversarial attacks, DSC models maintain near-baseline accuracy, offering a promising path toward more robust computer vision systems.
Introduction
Deep feed-forward convolutional networks (DCNs) have become the standard for image classification. However, their reliance on discriminative features makes them vulnerable to adversarial examples. These perturbations are typically high-frequency patterns that lack semantic meaning to humans but are statistically useful for a network's decision boundary. Because many DCNs learn similar idiosyncratic features, an attack designed for one model (e.g., DenseNet) often succeeds against others (e.g., ResNet or VGG). This phenomenon, known as transferability, poses a significant security risk.
This research explores whether moving from purely discriminative, feed-forward architectures to generative, recurrent models can mitigate these vulnerabilities. We focus on Deep Sparse Coding (DSC), a framework that seeks to represent data using a sparse set of learned basis functions, mimicking the processing found in the primary visual cortex.
Methods: Deep Sparse Coding Architecture
Unlike standard DCNs that process information in a single forward pass, the DSC architecture is a dynamical system. It functions as a generative autoencoder that attempts to reconstruct the input using a "dictionary" of basis vectors.
Key components of the DSC model include:
- Sparse Coding and LCA: The model employs the Locally Competitive Algorithm (LCA). Neurons compete to represent the input, and through lateral inhibition, only the most relevant neurons remain active, resulting in a sparse representation.
- Recurrent Dynamics: The network iterates until it settles into a stable state, or "attractor basin." This recurrence makes the model less sensitive to small input perturbations.
- Top-Down Feedback: Higher-level layers provide feedback to lower layers, ensuring that the low-level features are consistent with the global interpretation of the image.
- Classification: A simple linear Support Vector Machine (SVM) is trained on the stable internal representations produced by the DSC layers to perform final classification tasks.
Experimental Setup
The study utilized a binary classification task derived from the Labeled Faces in the Wild (LFW) dataset: identifying "Halle Berry" versus "Other People." The training set consisted of approximately 5,600 images. A holdout evaluation set of 150 images was subjected to four conditions:
- Original: Clean, unaltered images.
- Adversarial: Transferable examples generated using the Iterative Fast Gradient Sign Method (IFGSM) targeting a DenseNet121 model.
- Gaussian Noise: Images corrupted with random pixel-level noise.
- Low-pass Filtered: Images with high-frequency components removed via blurring.
Results and Discussion
The DSC model was compared against five prominent DCN architectures: ResNet50, InceptionV3, VGG16, MobileNetV2, and DenseNet121.
As shown in the results, the DCNs exhibited a sharp decline in accuracy when faced with transferable adversarial examples, with performance typically dropping from ~95% to the 70-80% range. DenseNet121, the white-box target, dropped to 0% accuracy. In contrast, the DSC model's performance remained remarkably stable at approximately 93-95% across all conditions, including adversarial perturbations, noise, and filtering.
Representation Stability
The robustness of DSC stems from its focus on generative, semantic features. The authors identified specific neurons (e.g., N-326) that responded to high-level facial features. These neurons remained active even when the input was heavily distorted by adversarial noise, whereas the internal representations of DCNs shifted significantly. The recurrent "attractor" nature of the DSC dynamics essentially "cleans" the input, pulling the representation back toward a known, valid state.
Conclusion and Future Work
Deep Sparse Coding offers a robust alternative to standard feed-forward networks for tasks where security and reliability are paramount. By prioritizing generative consistency and recurrent stability, DSC effectively ignores the non-semantic noise exploited by transferable adversarial attacks.
Future research will focus on the "white-box" vulnerability of DSC—whether an attacker with full knowledge of the recurrent dynamics can craft specialized perturbations. Additionally, scaling these architectures to more complex, multi-class datasets like ImageNet remains a key challenge due to the increased computational cost of recurrent iterations.
Summary updated