Training for Trustworthy Saliency Maps: Adversarial Training Meets Feature-Map Smoothing

Dipkamal Bhusal, Md Tanvirul Alam, Nidhi Rastogi

Abstract

Gradient-based saliency methods such as Vanilla Gradient (VG) and Integrated Gradients (IG) are widely used to explain image classifiers, yet the resulting maps are often noisy and unstable, limiting their usefulness in high-stakes settings. Most prior work improves explanations by modifying the attribution algorithm, leaving open how the training procedure shapes explanation quality. We take a training-centered view and first provide a curvature-based analysis linking attribution stability to how smoothly the input-gradient field varies locally. Guided by this connection, we study adversarial training and identify a consistent trade-off: it yields sparser and more input-stable saliency maps, but can degrade output-side stability, causing explanations to change even when predictions remain unchanged and logits vary only slightly. To mitigate this, we propose augmenting adversarial training with a lightweight feature-map smoothing block that applies a differentiable Gaussian filter in an intermediate layer. Across FMNIST, CIFAR-10, and ImageNette, our method preserves the sparsity benefits of adversarial training while improving both input-side stability and output-side stability. A human study with 65 participants further shows that smoothed adversarial saliency maps are perceived as more sufficient and trustworthy.

Introduction: The Reliability Crisis in Saliency Maps

Saliency maps—heatmaps highlighting the pixels a model prioritizes for a prediction—are essential for interpretability but notoriously unreliable. Standard methods like Vanilla Gradient (VG) often produce "noisy" visualizations that fluctuate wildly with minor input perturbations, even when the model's prediction remains constant. While previous research has focused on post-hoc algorithmic fixes, this paper investigates how the training procedure itself determines explanation quality.

The Stability Trade-off in Adversarial Training

The authors identify a critical "stability trade-off" inherent in adversarial training (AT). While AT is effective at producing cleaner, more focused (sparse) saliency maps by forcing the model to ignore non-robust features, it can inadvertently reduce "output-side stability." In these cases, a minor shift in model confidence (logits) can cause the explanation to change drastically, undermining human trust in the model's rationale.

Proposed Method: Feature-Map Smoothing (FMS)

To address this, the researchers propose a training-centered approach rooted in the mathematical relationship between model curvature and explanation stability. They demonstrate that a rapidly varying gradient field (high curvature) leads to unstable saliency maps.

Their solution combines Adversarial Training with a novel Feature-Map Smoothing (FMS) block. This "SmoothBlock" is inserted after the initial convolutional or residual layers of the network.

The SmoothBlock architecture integrates a differentiable Gaussian filter to stabilize internal feature maps.

As shown in Figure 3, the SmoothBlock applies a differentiable Gaussian filter to intermediate feature maps, followed by a 1x1 convolution and a residual connection. This acts as a low-pass filter, suppressing high-frequency fluctuations within the network's internal activations and smoothing the resulting gradient field.

Experimental Results

The approach was evaluated on FMNIST, CIFAR-10, and ImageNette using architectures including LeNet, ResNet-18, and WideResNet.

Quantitative Stability and Sparsity

The "Adversarial + Smoothing" models significantly improved Relative Output Stability (ROS) across all datasets compared to standard adversarial models. Crucially, this smoothing did not "blur away" the sparsity benefits of AT; the maps remained focused on discriminative features rather than background noise.

Visual Quality and Human Trust

Visual comparison of saliency maps across training methods, demonstrating the sparsity and clarity of the proposed approach. Source, p. 2

Visual comparison of saliency maps across training methods, demonstrating the sparsity and clarity of the proposed approach.

Visual inspection (Figure 1) confirms that smoothed adversarial maps are cleaner and more intuitive. A human study involving 65 participants validated these findings: 58% of participants preferred the smoothed adversarial maps, citing their ability to highlight important features without excessive noise, compared to only 27% for standard adversarial training.

Conclusion

This research demonstrates that explanation quality is a fundamental byproduct of a model's internal geometry. By reducing curvature through feature-map smoothing and robust training, we can produce saliency maps that are both sparse and stable, providing a more reliable foundation for human-AI trust. Future work will explore the scalability of these smoothing blocks to larger architectures like Vision Transformers and Large Language Models.

Summary updated

Abstract

Gradient-based saliency methods such as Vanilla Gradient (VG) and Integrated Gradients (IG) are widely used to explain image classifiers, yet the resulting maps are often noisy and unstable, limiting their usefulness in high-stakes settings. Most prior work improves explanations by modifying the attribution algorithm, leaving open how the training procedure shapes explanation quality. We take a training-centered view and first provide a curvature-based analysis linking attribution stability to how smoothly the input-gradient field varies locally. Guided by this connection, we study adversarial training and identify a consistent trade-off: it yields sparser and more input-stable saliency maps, but can degrade output-side stability, causing explanations to change even when predictions remain unchanged and logits vary only slightly. To mitigate this, we propose augmenting adversarial training with a lightweight feature-map smoothing block that applies a differentiable Gaussian filter in an intermediate layer. Across FMNIST, CIFAR-10, and ImageNette, our method preserves the sparsity benefits of adversarial training while improving both input-side stability and output-side stability. A human study with 65 participants further shows that smoothed adversarial saliency maps are perceived as more sufficient and trustworthy. Overall, our results demonstrate that explanation quality is critically shaped by training, and that simple smoothing with robust training provides a practical path toward saliency maps that are both sparse and stable.
BibTeX
@misc{bhusal2026training,
  author = {Dipkamal Bhusal and Md Tanvirul Alam and Nidhi Rastogi},
  title = {{Training for Trustworthy Saliency Maps: Adversarial Training Meets Feature-Map Smoothing}},
  year = {2026},
  eprint = {2603.07302},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.07302},
}

Loading related content...

Figure