Introduction
Deep Neural Networks (DNNs) have become the standard for semantic segmentation tasks, such as labeling pixels in autonomous driving scenes. However, these models are highly susceptible to adversarial examples (AEs)—carefully crafted inputs with imperceptible perturbations that lead to incorrect, high-confidence predictions. In safety-critical domains like autonomous driving, a static perception system can be "blinded" if an attacker gains knowledge of its architecture and weights. This paper investigates a redundant framework designed to enhance robustness by creating a "moving target" defense.
Methodology: The Redundant T-S-A Framework
The authors propose a Redundant Teacher-Student-Adaptive (T-S-A) framework consisting of three distinct models:
- Static Teacher ($\mathfrak{F}_T$): A high-capacity, pre-trained model.
- Static Student ($\mathfrak{F}_S$): A more efficient model trained to replicate the teacher's performance.
- Adaptive Student ($\mathfrak{F}_A$): A model initialized from the static student that undergoes continuous online adaptation using unlabeled data.
The primary innovation is the Inverse Feature Matching (IFM) loss. While standard knowledge distillation encourages students to mimic teacher features, the IFM loss penalizes the adaptive student if its internal representations are too similar to those of the vulnerable static student. This forces the adaptive student to find alternative feature paths to the correct output, effectively moving it out of the "vulnerability subspace" targeted by static attacks.
Experimental Evaluation
The framework was evaluated using the Cityscapes dataset, utilizing 2,950 finely-annotated images for initial training and 19,998 coarsely-annotated images (with labels removed) for online adaptation. Robustness was measured using the mean Intersection over Union (mIoU) ratio, comparing performance on adversarial images versus clean images.
Key Results
The inclusion of IFM loss significantly improved the resilience of the adaptive student against adversarial attacks. As shown in the results table below, without the IFM defense, the adaptive student's performance dropped to 35.14% of its clean-image accuracy under a moderate attack (S-AE, $\epsilon=1$). With the IFM defense enabled, the model maintained 80.51% of its original accuracy.
Discussion and Conclusion
The study demonstrates that intentional divergence from a known vulnerable model, guided by a high-performance teacher, is a viable defensive strategy. By constantly adapting on a stream of unlabeled data, the system presents a "moving target" that renders static adversarial maps obsolete.
While effective, this approach introduces a robustness-accuracy trade-off; the adaptive student's clean-image performance (approx. 53% mIoU) is lower than the static teacher's (approx. 75% mIoU). Future work will explore more sophisticated fusion strategies, such as majority voting among the three models, to further improve both base accuracy and overall system reliability in hostile environments.
Summary updated