Abstract
Multi-agent collaborative (MAC) perception significantly enhances the capabilities of autonomous vehicles and infrastructure by allowing them to share sensor data, such as LiDAR feature maps, to overcome individual limitations like occlusions. However, this connectivity introduces a critical security vulnerability: a single compromised agent can transmit "poisoned" data to degrade the entire network's perception. We present MADE (Malicious Agent Detection), a reactive defense framework that identifies and excludes malicious actors through statistical consistency checks. By evaluating discrepancies at both the feature and output levels, MADE restores perception performance to near-optimal levels without requiring ground truth labels or high communication overhead.
Introduction
In the context of autonomous driving, "two heads are better than one." Multi-agent collaboration allows vehicles to see around corners and through obstacles by aggregating data from neighboring agents. Yet, this collaboration is a double-edged sword. Traditional defenses, such as adversarial training, struggle with the high-dimensional complexity of shared feature maps. Furthermore, in real-world driving, an agent lacks access to "ground truth" to verify a neighbor's claims. The challenge lies in detecting malicious entities in real-time using only local information and previously collected benign data.
Methodology: The MADE Framework
MADE is a semi-supervised anomaly detector that treats malicious agent detection as a statistical consistency check. The core insight is that a liar’s story will eventually contradict the physical reality captured by the ego agent's own sensors. When an ego agent receives data, it performs two primary hypothesis tests:
1. Match Loss (ML) Statistic
This test checks for consistency at the output level. The ego agent compares the object detection proposals (bounding boxes) it generates in isolation against the proposals generated when collaborating with a neighbor. If the neighbor's data causes a drastic, illogical shift in these detections, the Match Loss spikes, signaling a potential attack.
2. Collaborative Reconstruction Loss (CRL) Statistic
This test checks for consistency at the feature level. The authors utilize an autoencoder trained on the "residuals"—the difference between ego-only features and benign collaborative features. If a neighbor sends a malicious, optimized perturbation, the autoencoder fails to reconstruct that residual accurately, resulting in a high CRL.
Statistical Decision Making
To make a final determination, MADE converts these statistics into conformal p-values and applies the Benjamini-Hochberg (BH) procedure. This statistical rigor ensures the system maintains strict control over false positives (wrongly accusing a benign agent) while maximizing the power to detect actual threats.
Experimental Evaluation
The researchers evaluated MADE using two major benchmarks:
- V2X-SIM: A simulated vehicle-to-everything dataset providing a controlled environment for multi-agent scenarios.
- DAIR-V2X: A large-scale, real-world dataset featuring cooperative perception between vehicles and roadside infrastructure.
Key Results
MADE demonstrated impressive robustness across both simulated and real-world environments:
- Performance Recovery: Under heavy adversarial attacks, MADE recovered 22.13% Average Precision (AP) on V2X-SIM and 22.60% on DAIR-V2X compared to a baseline with no defense.
- Near-Oracle Accuracy: MADE’s performance was nearly identical to an "Oracle" defender (a theoretical model with perfect knowledge of which agents are malicious), trailing by only 1.27% and 0.28% AP, respectively.
- Superior Detection: The framework significantly outperformed previous state-of-the-art methods like ROBOSAC, particularly in maintaining a high True Positive Rate while keeping the False Positive Rate at a target 5% level.
Discussion and Limitations
MADE represents a shift from proactive to reactive defense. Instead of attempting to make a model robust to all possible noise, it identifies the source of the attack and removes it from the network entirely. However, some limitations remain:
- The "Benign Ego" Assumption: The framework assumes the agent running the detector is itself benign. If the ego agent is compromised, the system cannot self-correct.
- Computational Overhead: Running an autoencoder and matching algorithm locally requires computational resources. Optimizing these for ultra-low latency in high-speed driving scenarios is a necessary next step.
- Adaptive Attacks: Future work will investigate "stealthy" adaptive attacks specifically designed to bypass both the ML and CRL statistics simultaneously.
Conclusion
MADE provides a robust, statistically grounded solution for securing multi-agent collaborative perception. By leveraging local sensor data as a proxy for truth, it effectively "unfriends" malicious actors, ensuring that the benefits of collaboration do not come at the cost of safety.
Summary updated