Abstract
Deep neural networks (DNNs) used in autonomous vehicles (AVs) are often fine-tuned "online" to adapt to new environments. This research identifies a critical vulnerability in this process: the "Bait and Switch" attack. By manipulating the physical environment with electronic billboards that create spurious correlations with traffic signals, an adversary can poison a model's decision-making logic. This "clean-label" attack requires no digital access to the training pipeline, yet it can cause catastrophic misclassifications during deployment.
Introduction
Autonomous driving systems rely on DNNs that are notoriously "lazy" learners, often latching onto the most prominent statistical correlations in a dataset rather than identifying true causal relationships. To handle local variations in infrastructure, these models are frequently fine-tuned as they enter new cities. While traditional data poisoning attacks (backdoors or Trojans) typically require an attacker to inject malicious code or labels into a digital pipeline, this paper explores a more insidious threat: physical environmental poisoning.
The Bait and Switch Framework
The researchers propose a two-stage framework designed to exploit the gap between correlation and causality in machine learning models.
1. The Bait Stage
During the online learning phase in a new environment (e.g., "City B"), the attacker places electronic billboards near traffic intersections. These billboards display specific high-contrast images—such as a dog, cat, or fish—that are perfectly synchronized with the traffic light states. Because the DNN finds these large billboard images easier to identify than the small, distant traffic light bulbs, it learns to associate the "bait" image (e.g., a dog) with a specific state (e.g., "Green").
2. The Switch Stage
Once the model has been poisoned and deployed, the attacker "switches" the logic. By displaying the "Green-associated" image on the billboard while the actual traffic light is Red, the attacker tricks the vehicle into a dangerous misclassification.
Experimental Methodology
The study utilized the CARLA simulator, an Unreal Engine 4-based environment, to generate high-fidelity training and testing data. The researchers employed a ResNet18 architecture modified for a 3-class traffic light classification task (Red, Yellow, Green).
- Town A (Clean): 10,400 images per state for initial training.
- Town B (Poisoned): 12,400 images per state for fine-tuning, where varying numbers of intersections (from 3 to 37) were paired with adversarial billboards.
Key Insights
- Physical Clean-Label Attack: This is a "clean-label" attack because the human expert providing the ground truth labels sees the actual traffic light and labels it correctly. The poison exists entirely in the input pixels of the environment, making it invisible to standard data auditing.
- Generalization of Spurious Concepts: The model does not just learn the poison at specific intersections. Once the "Dog = Green" association is formed, the model applies this logic to any intersection where a billboard is present, regardless of whether it encountered that specific location during training.
- Poisoning Efficiency: The attack is highly efficient; poisoning as few as 10% of the traffic lights in a city was sufficient to significantly degrade system reliability.
Results
The impact of the attack was severe. While a cleanly fine-tuned model achieved 98.25% accuracy, the poisoned model's accuracy plummeted to 33.89% when the switch was activated across all intersections. Even with partial poisoning (only 3 out of 37 lights), accuracy at those specific locations dropped to 77%. Furthermore, the attack remained effective even when only the final layers of the network were retrained, suggesting that "freezing" early layers is an insufficient defense.
Conclusion
The "Bait and Switch" attack demonstrates that for an autonomous system, the entire physical world is a potential attack vector. Future work must focus on developing architectures that are robust to non-causal features, perhaps by forcing models to ignore background context or using uncertainty-based detection to identify when a model is relying on suspicious environmental cues. As AVs become more prevalent, securing the physical training environment is as critical as securing the digital one.
Summary updated