Abstract
Neural networks are increasingly deployed in safety-critical domains where robustness to environmental shifts is paramount. While defense against pixel-level white noise is well-studied, real-world perturbations often exhibit spatial structure, such as bias fields in medical imaging or lighting variations in autonomous driving. This paper introduces a novel training framework, ROBNET, utilizing the RSIP-SSIP algorithm to provide certified robustness against these structured intensity shifts. By replacing loose Interval Bound Propagation (IBP) with Standard Symbolic Interval Propagation (SSIP), our method achieves up to a 31% increase in certified robustness over state-of-the-art baselines while maintaining high standard accuracy.
Introduction
Standard adversarial robustness research typically focuses on independent pixel-level perturbations. However, many practical distribution shifts are structured. A "bias field" represents smooth, spatially varying intensity changes across an image. In MRI scans, these appear as hardware-induced artifacts; in natural images, they model shadows or shifting illumination.
Existing "certifiably robust" training methods, which provide mathematical guarantees on model output stability, often fail when applied to bias fields. Techniques like Interval Bound Propagation (IBP) produce overly conservative (loose) bounds because they ignore the inherent dependencies between pixels in a smooth gradient. This leads to poor optimization and low certified accuracy.
Methodology: The RSIP-SSIP Framework
To address these challenges, we propose a two-pronged approach implemented in the ROBNET toolkit.
Problem Reduction via Transform Networks
We first demonstrate that bias field robustness can be mathematically reduced to a white-noise robustness problem. We prepend a Bias Field Transform Network layer to the target model. This layer maps low-dimensional bias field coefficients (e.g., polynomial coefficients) to the full-resolution perturbed image, allowing us to analyze the perturbation in a compact parameter space.
The RSIP-SSIP Algorithm
The core of our training method is the RSIP-SSIP algorithm. Traditional robust training often pairs Reversed Symbolic Interval Propagation (RSIP) with IBP. We prove that IBP is fundamentally ill-suited for bias fields as it treats every pixel as independent, discarding the structural information of the smooth field.
We replace IBP with Standard Symbolic Interval Propagation (SSIP). SSIP maintains linear equations to track how input coefficients propagate through hidden layers. This results in significantly tighter mathematical bounds, providing a more accurate "robustness loss" signal for the optimizer.
Experimental Evaluation
We evaluated RSIP-SSIP on two benchmarks: CIFAR10 (natural images) and an MRI Brain Tumor Dataset (medical imaging).
Key Results
Our experiments demonstrate that RSIP-SSIP consistently outperforms existing methods in both robustness and accuracy.
- Certified Robustness Gains: On CIFAR10, RSIP-SSIP achieved up to 31% higher certified robustness than the previous state-of-the-art (RSIP-IBP).
- Performance Stability: Unlike many robust training methods that trade off standard accuracy for safety, RSIP-SSIP maintained high test accuracy even as the perturbation radius (\epsilon) increased.
- Medical Imaging Efficacy: In the MRI task, our method achieved 96.6% certified robustness against significant bias field artifacts, whereas standard models collapsed to 0% robustness.
Discussion and Future Work
The success of RSIP-SSIP stems from the low-dimensionality of bias fields. Because these fields are defined by a small number of parameters, symbolic methods like SSIP—which are typically computationally prohibitive for high-dimensional inputs—become feasible.
While highly effective for intensity-based shifts, future research will explore scaling these symbolic propagation techniques to larger Transformer-based architectures and extending the framework to handle geometric perturbations like rotations and scaling.
Summary updated