Abstract
Aerial object detection systems powered by Deep Neural Networks (DNNs) are critical for modern surveillance, yet they remain vulnerable to adversarial attacks. This paper introduces the Adaptive-Patch-based Physical Attack (AP-PA) framework, a novel approach designed to generate robust, scale-adaptive adversarial patches. By utilizing a unique "Mean Score" loss function and physical adaptation techniques, AP-PA achieves significant performance degradation in both digital and physical environments. Notably, the framework demonstrates that patches can effectively "cloak" targets even when placed adjacent to them, rather than directly on the object surface.
Introduction
Deep Neural Networks (DNNs) serve as the backbone for identifying civilian and military hardware in aerial and satellite imagery. Despite their sophistication, these networks possess a critical vulnerability: adversarial examples. By introducing mathematically optimized patterns—adversarial patches—into a scene, an attacker can effectively render objects invisible to detection algorithms.
Existing adversarial patches often struggle in aerial contexts due to three primary factors: environmental sensitivity (lighting and angle changes), computational inefficiency during generation, and the physical constraint of needing to be applied directly to the target. The AP-PA framework addresses these hurdles while establishing a rigorous benchmark for evaluating aerial detection security.
Methodology
The AP-PA framework incorporates several engineering innovations to ensure the effectiveness of attacks in the physical world:
Physical Adaptation (PT)
To bridge the gap between digital simulation and physical reality, the algorithm incorporates a transformation layer during training. This layer simulates real-world distortions, including digital noise, varying scales, rotations, and shifting lighting conditions. This ensures that a printed patch remains effective under varying solar angles or camera altitudes.
Scale-Adaptive Positioning
Aerial targets vary significantly in size depending on the sensor's altitude. AP-PA automatically scales the patch and adjusts its distance from the target based on the object's bounding box. This ensures the adversarial "distraction" remains proportional to the target regardless of the drone's height.
The "Mean Score" Loss Function
Traditional adversarial attacks typically focus on minimizing the confidence score of the single most prominent detection. AP-PA introduces a new objective loss ($L_{obj}$) that considers the mean objectiveness score of all detected objects in a frame. This "shotgun" approach suppresses multiple potential detections simultaneously, leading to more stable optimization and faster convergence compared to state-of-the-art methods.
Printability Constraints
The framework includes a Non-Printability Score ($L_{nps}$) to ensure that the colors generated during optimization fall within the reproducible gamut of standard CMYK printers, facilitating real-world deployment.
Experimental Setup
The researchers utilized two primary datasets for training and validation:
- DOTA: A comprehensive dataset for aerial object detection containing nearly 190,000 instances across 15 categories, used to train the "victim" detectors.
- RSOD: A specialized dataset containing aircraft, used to optimize and refine the adversarial patches.
Key Insights and Results
Distraction vs. Destruction
A pivotal finding of this research is that a patch does not need to physically cover a target to be effective. By placing a patch outside the object's boundaries, the attacker exploits the way DNNs process "objectiveness." The patch essentially siphons the detector's attention, causing the model to ignore the actual target entirely.
Transferability and Efficiency
The study found that detectors with similar architectures (e.g., various versions of YOLO) are highly susceptible to the same patches. Furthermore, patches generated against simpler models like YOLOv2 demonstrated high transferability, successfully fooling more complex Transformer-based detectors. The transition to a "mean-score" loss function also provided a significant efficiency boost, preventing optimization from stalling on single detections.
Performance Metrics
The results indicate a severe vulnerability in current aerial systems:
- White-Box Settings: When the attacker has full knowledge of the model, the Average Precision (AP) of detectors dropped by up to 87.86%.
- Black-Box Settings: Even with unknown models, the AP drop reached 85.48%.
- Physical Validation: Scaled physical experiments using aircraft models confirmed that these patches successfully hide targets in real-time, maintaining effectiveness across varying camera angles.
Conclusion and Future Work
AP-PA demonstrates that adversarial patches are a viable threat to aerial surveillance, even when they are not applied directly to the target. While the framework is highly effective, the current patches are visually conspicuous to human observers. Future research will focus on "stealthy" adversarial patterns that blend into natural textures, such as runway markings or roof patterns. Additionally, the development of adversarial training methods is necessary to bolster the defenses of security-critical aerial detection systems.
Summary updated