Abstract
Deep neural networks suffer from significant performance degradation when exposed to common corruptions such as noise, blur, weather, and digital distortions, limiting their reliability in real-world applications. In this paper, we propose AR2 (Attention-Guided Repair for Robustness), a simple yet effective method to enhance the corruption robustness of pretrained CNNs. AR2 operates by explicitly aligning the class activation maps (CAMs) between clean and corrupted images, encouraging the model to maintain consistent attention even under input perturbations. Our approach follows an iterative repair strategy that alternates between CAM-guided refinement and standard fine-tuning, without requiring architectural changes. Extensive experiments show that AR2 consistently outperforms existing state-of-the-art methods in restoring robustness on standard corruption benchmarks (CIFAR-10-C, CIFAR-100-C and ImageNet-C), achieving a favorable balance between accuracy on clean data and corruption robustness.
Introduction: The Vulnerability of the CNN "Gaze"
Deep Neural Networks (DNNs) are notoriously brittle. A model that perfectly identifies a "dog" in a clear studio photo might completely fail if that same photo is slightly blurred, grainy, or taken in the rain. These "common corruptions" represent a major hurdle for deploying AI in the real world, where sensors are imperfect and environments are unpredictable.
The authors observed a specific phenomenon: when a CNN encounters a corrupted image, its internal "attention" shifts. Instead of looking at the discriminative features of the object (like the ears or snout of a dog), the model's focus becomes fragmented or shifts toward the corruption patterns themselves. The challenge is to "repair" these models so they maintain high accuracy on clean data while becoming robust to noise, without requiring complex architectural changes.
Methodology: The AR2 Framework
The core idea of AR2 is Attention Alignment. The researchers propose a two-stage iterative process that alternates between refining the model's focus and restoring its classification accuracy.
1. CAM-Guided Refinement
The method uses Class Activation Maps (CAMs)—heatmaps that show which pixels a model "attends" to when making a prediction. AR2 uses a frozen "reference model" (the original clean model) to provide a "ground truth" for where the attention should be. It then trains a "repairing model" to ensure its attention on both clean and corrupted images matches this reference. This is achieved using a Mean Squared Error (MSE) loss on the CAMs of the top-$k$ predicted classes. Crucially, AR2 updates only the convolutional backbone during this stage, keeping the decision-making layers frozen to maintain the model's original logic.
2. Standard Fine-tuning
Because the first stage focuses purely on attention and not classification accuracy, the second stage uses standard supervised learning (cross-entropy loss) on both clean and corrupted images to restore and refine the model's predictive power.
Key Insights: Attention as a Robustness Signal
The most significant finding is that fixing where a model looks is more effective than simply feeding it more augmented data. If the model learns to ignore the "noise" and focus on the "object" via CAM alignment, robustness follows naturally.
As shown in the visualization above, vanilla models often lose focus or attend to background noise when faced with corruptions. AR2 successfully "repairs" this gaze, ensuring the model remains focused on the relevant semantic features.
Experimental Results
The authors evaluated AR2 on three industry-standard robustness benchmarks: CIFAR-10-C, CIFAR-100-C, and ImageNet-C. These datasets contain 15 types of algorithmic corruptions across five severity levels.
AR2 consistently set new state-of-the-art (SOTA) marks for corruption robustness:
- CIFAR-10-C: Achieved a mean Corruption Error (mCE) of 30.4%, a massive 10.1% improvement over the next best method (AugMix).
- CIFAR-100-C: Reduced mCE to 48.7%, outperforming the baseline by 12.6%.
- ImageNet-C: Reduced mCE by 10.1% compared to AugMix, dominating in 12 out of 15 corruption categories.
- Clean Accuracy: Unlike many robustness techniques that "tax" the model's performance on perfect images, AR2 maintained clean accuracy levels very close to (and sometimes better than) existing SOTA methods.
Conclusion and Future Work
AR2 demonstrates that attention alignment is a powerful tool for enhancing CNN robustness. By decoupling attention repair from accuracy fine-tuning, the framework achieves significant gains in corruption resilience without sacrificing clean performance.
The primary limitation is the current per-corruption repair paradigm, which requires training separate models for specific corruption types. Future work aims to develop a unified AR2 variant that can jointly optimize for multiple corruptions simultaneously, creating a single, "all-weather" robust model.
Summary updated