Główne pojęcia
Bayesian trust estimation enhances security in multi-agent autonomy by mapping sensor data to trust pseudomeasurements.
Streszczenie
The content discusses the vulnerability of track scoring algorithms in multi-agent autonomy to adversarial attacks. It introduces a Bayesian trust estimation framework to enhance security by mapping sensor data to trust pseudomeasurements. The analysis includes two case studies with and without prior information on agent trust, showcasing the impact of prior knowledge on trust estimation outcomes.
I. Introduction
- Importance of collaborative sensor fusion in safety-critical environments.
- Need for security-awareness in multi-agent collaboration.
II. Multiple Target Tracking (MTT)
- Challenges of false positives and false negatives in object existence determination.
- Central tasks and algorithms for multiple target tracking.
III. Security Analysis of Track Scoring
- Vulnerability of track scoring to adversarial manipulation.
- Threat model considerations and analysis of track score updates.
IV. Estimation of Track and Agent Trust in MTT
- Bayesian approach to estimating trust using pseudomeasurements.
- Decomposition into subproblems for sequential updating.
V. Multi-Agent Trust Experiments
- Evaluation of proposed trust estimation models on two case studies.
- Impact of prior information on agent trust on the outcomes.
Statystyki
We prove that even when benign agents outnumber adversaries, attackers need only a small number of frames to establish high-confidence FP tracks that are mistakenly believed to be real objects.
Cytaty
"Track scoring is vulnerable to adversarial manipulation."
"Our approach estimates whether tracks and agents are trustworthy via hierarchical Bayesian updating."