Abstract: Correctly classifying adversarial examples is an essential but challenging requirement for safely deploying machine learning models. As reported in RobustBench, even the state-of-the-art adversarially trained models struggle to exceed 67% robust test accuracy on CIFAR-10, which is far from practical. A complementary way towards robustness is to introduce a rejection option, allowing the model to not return predictions on uncertain inputs, where confidence is a commonly used certainty proxy. Along with this routine, we find that confidence and a rectified confidence (R-Con) can form two coupled rejection metrics, which could provably distinguish wrongly classified inputs from correctly classified ones. This intriguing property sheds light on using coupling strategies to better detect and reject adversarial examples. We evaluate our rectified rejection (RR) module on CIFAR-10, CIFAR-10-C, and CIFAR-100 under several attacks including adaptive ones, and demonstrate that the RR module is compatible with different adversarial training frameworks on improving robustness, with little extra computation.
Introduction
Deep learning models are notoriously fragile when faced with adversarial examples: inputs specifically crafted to trigger misclassifications. While Adversarial Training (AT) helps, it hits a performance ceiling that is still far from practical for safety-critical systems. Even the best adversarially trained models struggle to reach 70% accuracy on standard benchmarks like CIFAR-10.
A common fallback is to give the model a "rejection option"—essentially allowing it to say "I don't know" instead of providing a wrong answer. However, existing rejection methods often fail because they overestimate certainty on misclassified samples, and the "rejector" itself is often just as vulnerable to adversarial attacks as the classifier it seeks to protect. This paper proposes a novel rejection mechanism called Rectified Rejection (RR) to address these challenges.
The Rectified Rejection Framework
The core innovation of this work is the introduction of True Confidence (T-Con) and its practical proxy, Rectified Confidence (R-Con).
T-Con and R-Con
- T-Con (The Oracle): This is defined as the probability the model assigns to the actual ground truth label ($f_\theta(x)[y]$). It serves as a perfect certainty metric but is an "oracle" value only available during training.
- R-Con (The Proxy): Since the true label is unknown at inference, the authors train an auxiliary module ($A_\phi$) to predict T-Con. R-Con is defined as the product of the model's predicted confidence and this auxiliary output.
Coupled Rejection Strategy
Instead of relying on a single threshold, the RR module uses a "coupling" strategy. First, inputs are filtered by a standard confidence threshold. For samples that pass this initial check, a second R-Con threshold is applied.
The authors provide a mathematical proof (Theorem 1) showing that if an input has high confidence but low R-Con, it is almost certainly misclassified. This coupling creates a gap that adversarial attacks find very difficult to bridge, as an attacker would need to simultaneously maximize both metrics for a misclassified sample.
Theoretical and Empirical Separability
A key insight of the paper is that learning to reject is fundamentally easier than learning a perfectly robust classifier. Through Theorem 2, the authors argue that learning a "$\xi$-error" rejector (one that approximates T-Con well) is a simpler task. This is visualized in the distribution of R-Con values across correctly and wrongly classified samples.
As shown in the figure above, wrongly classified points (red) tend to cluster at lower R-Con values compared to correctly classified ones (green). This property allows the RR module to reliably separate correct predictions from incorrect ones, even under heavy attack.
Experimental Evaluation
The researchers evaluated the RR module using standard computer vision benchmarks, including CIFAR-10, CIFAR-100, and CIFAR-10-C (to test general out-of-distribution robustness). The module was tested against a battery of attacks, including the formidable AutoAttack and custom adaptive attacks designed specifically to break the rejector.
Key Results
- Superior Robustness: On CIFAR-10, the RR module achieved a TPR-95 accuracy (accuracy when 95% of clean samples are accepted) of approximately 58% under PGD-100 attacks, outperforming strong baselines like CCAT and GDA.
- ROC-AUC Gains: The method consistently showed higher ROC-AUC scores compared to traditional methods like Kernel Density (KD) or Local Intrinsic Dimensionality (LID).
- Efficiency: Because the RR module shares the "backbone" feature-extracting layers with the classifier, it adds negligible computational overhead during both training and inference.
Conclusion
Rectified Rejection offers a promising path forward for adversarial robustness by shifting the focus from perfect classification to reliable rejection. By coupling standard confidence with a learned Rectified Confidence, the RR module creates a provable gap that distinguishes correct and incorrect classifications. While current limitations include the fact that this separability applies to a portion of the input space rather than all points, the framework is compatible with various adversarial training frameworks and provides a significant boost to practical robustness with minimal extra computation.
Summary updated