MAIVA: Multi-Agent Integrity Verification Architecture

National Importance

Multi-agent autonomous systems (drone swarms, robotic teams, distributed sensor networks) present unique governance challenges. Unlike single-agent systems where trust evaluation is internal, multi-agent systems must aggregate trust across agents that may be compromised, faulty, or adversarially controlled. The Byzantine Generals Problem (Lamport, Shostak & Pease, 1982) formalizes this challenge: how can agents reach consensus when some agents may be unreliable?

Current swarm architectures typically assume reliable participation from all agents, creating vulnerability to compromised agents that can corrupt collective decisions. MAIVA introduces trust-conditioned participation where each agent's role in the mission is continuously evaluated and dynamically constrained based on its computed trust and authority state.

MAIVA Architecture

MAIVA extends HMAA authority governance from single-agent to multi-agent environments. Each agent executes a local SATA-HMAA-CARA governance stack and transmits signed trust reports to a Mission Authority Node (MAN). The MAN aggregates trust using algorithms with the Byzantine bound as its design basis:

Trimmed Weighted Median Aggregation: roster sized to the 3f+1 Byzantine bound as a configuration reference
Three-layer CUSUM-augmented anomaly detection
Graduated escalation with per-level action permissions
DoDD 3000.09 action gate classification

MAIVA Pipeline

Key Contributions

• Extension of HMAA authority governance from single-agent to multi-agent environments
• Trimmed weighted median aggregation, with the roster sized to the 3f+1 Byzantine bound as a configuration reference
• Three-layer CUSUM-augmented anomaly detection for trust drift identification
• Graduated escalation with per-level action permissions (Observe, Report, Constrain, Isolate)
• DoDD 3000.09 action gate classification for mission-critical decisions
• Trust-conditioned participation: compromised agents automatically isolated
• Swarm reconfiguration protocol redistributing tasks among trusted agents
• 37 self-tests and TLA+ formal specification

Role in the Governance Stack

MAIVA extends the single-agent SATA-HMAA-CARA pipeline to coordinated multi-agent environments. Each agent runs the full local governance stack independently, then MAIVA aggregates trust reports at the swarm level. This dual-layer architecture is implemented in the UAV swarm governance extension, where each drone's participation is conditioned on its computed trust and authority state. MAIVA integrates with ADARA for detecting adversarial agents attempting to manipulate swarm consensus.

The Multi-Agent Trust Problem

Single-agent governance (SATA-HMAA-CARA) assumes the autonomous system is evaluating its own sensors. In multi-agent systems (drone swarms, robotic teams, distributed sensor networks), a fundamentally different challenge arises: each agent must trust not only its own sensors but also the reports from other agents, some of which may be compromised.

Lamport, Shostak, and Pease (1982) formalized this as the Byzantine Generals Problem: how can distributed agents reach consensus when some agents may be faulty or adversarial? In autonomous systems, a compromised drone in a swarm could transmit false trust reports, corrupting collective decisions and potentially causing the entire swarm to take unsafe actions based on manipulated consensus.

Current swarm architectures (Brambilla et al., 2013) typically assume reliable participation and focus on coordination algorithms (formation control, task allocation). MAIVA introduces trust-conditioned participation where each agent's ability to influence swarm decisions is dynamically gated by its computed trust and authority state.

Trust Aggregation (Byzantine Bound as Design Basis)

Trimmed Weighted Median

MAIVA aggregates per-agent trust reports using a trimmed weighted median, with the roster sized to the 3f+1 Byzantine bound as a configuration reference. The trimming removes the most extreme trust reports (potential adversaries) before computing the weighted median of remaining reports. This bounds the influence of up to f adversarial reports on the aggregate; the bound is used as a configuration reference, not as a proof of correct aggregation. See Formal Properties.

Three-Layer Anomaly Detection

Graduated Escalation

When an agent is flagged as potentially compromised, MAIVA applies graduated escalation rather than immediate isolation:

MAIVA Swarm Trust Simulation

The MAIVA simulation demonstrates multi-agent trust aggregation with configurable swarm size, adversary count, and attack strategies. Users can observe how the trimmed-median aggregation is intended to bound the influence of compromised agents on swarm decisions.

API Implementation

POST /swarm/aggregate

{
  "roster_size": 7,
  "agent_reports": [
    {"id": "drone-1", "trust": 0.91, "auth": "A3"},
    {"id": "drone-2", "trust": 0.88, "auth": "A3"},
    {"id": "drone-3", "trust": 0.12, "auth": "A0"},
    {"id": "drone-4", "trust": 0.85, "auth": "A2"},
    {"id": "drone-5", "trust": 0.90, "auth": "A3"},
    {"id": "drone-6", "trust": 0.87, "auth": "A3"},
    {"id": "drone-7", "trust": 0.82, "auth": "A2"}
  ],
  "max_adversaries": 2
}

{
  "consensus_trust": 0.87,
  "mission_authority": "A3",
  "isolated_agents": ["drone-3"],
  "isolation_reason": "trust below threshold",
  "escalation_level": "CONSTRAIN",
  "active_roster": 6,
  "cusum_alerts": ["drone-3"],
  "task_redistribution": {
    "drone-3_tasks": "reassigned to drone-1, drone-5"
  }
}

Byzantine Resilience Formula

Minimum roster for f adversaries: n ≥ 3f + 1
7 agents → sized for f = 2 (3×2+1 = 7)
12 agents → sized for f = 3 (3×3+1 = 10 ≤ 12)

Selected References

• Lamport, L., Shostak, R. & Pease, M. (1982). The Byzantine Generals Problem. ACM TOPLAS, 4(3), 382-401.
• Brambilla, M. et al. (2013). Swarm Robotics: A Review from the Swarm Engineering Perspective. Swarm Intelligence, 7(1), 1-41.
• Olfati-Saber, R., Fax, J.A. & Murray, R.M. (2007). Consensus and Cooperation in Networked Multi-Agent Systems. Proc. IEEE, 95(1), 215-233.
• U.S. DoD (2023). Directive 3000.09: Autonomy in Weapon Systems.
• Cho, J.-H. et al. (2015). A Survey on Trust Modeling. ACM Computing Surveys, 48(2), Article 28.

Formal Design Basis and Open Verification Status

Known Limitations and Failure Modes

Simulation Reproducibility

The simulation supports single-architecture mode (MAIVA consensus only) and full pipeline mode (MAIVA integrated with SATA, HMAA, ADARA, FLAME, and CARA). Both configurations demonstrate MAIVA behavior under Byzantine fault conditions.

Cite This Work

If you reference this architecture in your research, please use one of the following citation formats:

@misc{oktenli2026maiva,
  author       = {Oktenli, Burak},
  title        = {Multi-Agent Integrity Verification Architecture},
  year         = {2026},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.19015517},
  url          = {https://doi.org/10.5281/zenodo.19015517},
  note         = {Georgetown University}
}

About This Project

This architecture is part of the authority-governed autonomy research program by Burak Oktenli at Georgetown University (M.P.S. Applied Intelligence, in progress). It is published on Zenodo with DOI 10.5281/zenodo.19015517 under CC BY 4.0.

Related: Full Research Portfolio · All Repositories · Rover Testbed · UAV Platform · Evaluation Protocol