Abstract
Standard multi-target tracking (MTT) algorithms are surprisingly easy to fool; even a single compromised agent can trick a majority of benign agents into "confirming" a fake object. This paper introduces a hierarchical Bayesian framework that estimates the trustworthiness of both tracks and agents in real-time, effectively filtering out adversarial data in collaborative sensing environments.
Introduction
In collaborative autonomy—such as fleets of autonomous vehicles or drone swarms—agents share sensor data to construct a joint representation of their environment. Multi-Target Tracking (MTT) is the standard approach for managing this data. To prevent "ghost objects" or false positives, MTT systems typically employ a "track score" to distinguish real objects from sensor noise.
However, these systems lack inherent security against coordinated attacks. An adversary infiltrating the network can inject "time-correlated" false positives—fake detections designed to mimic moving objects. The authors prove that traditional track scores are highly vulnerable: an attacker needs only a few frames to establish a high-confidence fake track, even if numerous benign agents report no detection.
Methodology: Hierarchical Bayesian Trust Estimation
The proposed solution is a two-pronged approach involving vulnerability analysis and a novel trust estimation framework.
Vulnerability Analysis
The authors derive Theorem 1, calculating the minimum number of frames ($T_{min}$) an adversary requires to force a system to confirm a fake track. This provides a mathematical bound on the speed of potential system compromise.
Trust Estimation Framework
The system shifts from tracking only objects to tracking "trust" simultaneously.
- Trust Pseudomeasurements (PSMs): In the absence of direct trust sensors, the system generates "pseudomeasurements." If an agent reports an object that neighboring agents with overlapping fields of view do not see, a low-trust PSM is generated for both the track and the reporting agent.
- Hierarchical Bayesian Updating: Trust is modeled as a probability distribution using Beta-Bernoulli conjugate pairs. The system employs an alternating update process: current agent trust informs track trust updates, which in turn refine agent trust scores.
- Prior Incorporation: The framework can integrate "priors." For instance, if a specific platform is known to be high-security and reliable, this certainty helps resolve ambiguities in conflicting reports from other agents.
Experimental Evaluation
The framework was evaluated using simulated multi-agent surveillance scenarios involving three agents with partially overlapping circular fields of view. Two primary adversarial cases were tested:
- Case 1: Two agents are partially compromised and provide malicious false positives.
- Case 2: A single agent provides multiple malicious false positives.
Key Results
- Theorem 1 Validation: Simulations confirmed that with typical parameters (e.g., $P_D = 0.9$), an adversary can confirm a fake track in just 3 to 6 frames, even when outnumbered by benign agents.
- Successful Distrust: The Bayesian model successfully drove trust scores of adversarial tracks toward zero while maintaining high trust for legitimate objects.
- Rapid Convergence: The hierarchical updates demonstrated rapid convergence in practice, making the framework suitable for real-time autonomous applications.
Discussion and Insights
A critical finding is the "Majority" Fallacy: a majority of benign agents does not guarantee safety. Because "hits" (detections) often carry more mathematical weight than "misses," a single attacker can "out-gain" the collective silence of the majority.
Furthermore, observability is essential; trust can only be estimated within overlapping fields of view. If an object is visible to only one agent, the system cannot mathematically distinguish between a malicious report and a unique line of sight. In such cases, priors become vital for resolving ambiguity.
Limitations and Future Work
The current model faces the "Lone Wolf" problem, where it cannot verify tracks visible to only one agent. Future work may involve dynamic sensor resource management to re-task agents for verification. Additionally, the authors plan to address "state estimation attacks" and investigate computational scaling for large-scale deployments in dense environments.
Summary updated