SATA: Sensor Attestation and Trust Anchoring

National Importance

Sensor spoofing and degradation represent primary attack vectors against autonomous systems. GPS spoofing can redirect unmanned vehicles while the navigation system maintains full confidence. LiDAR interference can create phantom obstacles or blind perception systems. IMU manipulation through vibration or electromagnetic interference corrupts orientation data. These attacks exploit a fundamental weakness: current systems lack formal mechanisms to evaluate and quantify trust in their own sensor data.

The NIST AI Risk Management Framework identifies measurement and monitoring of AI system inputs as essential governance practices. Khaleghi et al. (2013) survey multisensor data fusion and identify trust-aware fusion as an open research challenge. SATA addresses this by providing continuous, mathematically grounded sensor trust evaluation using Dempster-Shafer evidence theory.

SATA Architecture

SATA computes a continuous trust scalar τ ∈ [0,1] for each sensor using weighted Dempster-Shafer belief functions over a binary frame of discernment Θ = {Trusted, Compromised}. Per-sensor basic probability assignments (BPAs) are constructed from four diagnostic components:

m_i({Trusted}) = τ(s_i, t) × w_i
m_i({Compromised}) = (1 - τ(s_i, t)) × w_i
m_i(Θ) = 1 - w_i

Combination: (m_1 ⊕ m_2)(A) = (1/K) × Σ[m_1(B) × m_2(C)] for B∩C=A

SATA Pipeline

Key Contributions

• Continuous sensor trust evaluation using weighted Dempster-Shafer belief functions
• Binary frame of discernment Θ = {Trusted, Compromised} for each sensor
• Four diagnostic components: internal consistency, cross-sensor agreement, temporal stability, physical plausibility
• Cross-sensor validation with disagreement penalties (trust × 0.3 reduction)
• Asymmetric trust dynamics: fast decay (0.5s) and slow recovery (5.0s)
• Single-sensor veto capability: fused trust drops by 0.3 or more
• Foundation layer for HMAA authority computation across all platforms
• Validated in rover testbed (5 sensors), UAV platform (8 sensors), and all four BLADE platforms

Role in the Governance Stack

SATA is the foundation layer of the governance stack. It provides the fused trust scalar τ that drives all downstream authority decisions. HMAA uses τ to compute authority levels. ADARA modifies SATA's output by incorporating adversarial deception probability. In multi-agent systems, per-agent SATA trust feeds into MAIVA for swarm-level aggregation. Both the rover testbed (5 sensors) and UAV platform (8 sensors) implement SATA as their primary trust evaluation mechanism.

The Sensor Trust Problem

Autonomous systems rely entirely on sensor data to perceive their environment and make operational decisions. When sensor data is unreliable, whether due to hardware degradation, environmental interference, or adversarial manipulation, every downstream decision is compromised. The fundamental challenge is: how does an autonomous system know whether to trust its own sensors?

Current sensor fusion approaches (Khaleghi et al., 2013) typically assume all sensors are reliable and focus on optimally combining their outputs. When a sensor fails, most systems either ignore it entirely or use simple voting schemes. These approaches lack the mathematical rigor needed for safety-critical governance: they cannot express partial trust, cannot detect sophisticated spoofing, and cannot quantify their own confidence in sensor integrity.

GPS spoofing demonstrations have shown that autonomous vehicles can be redirected while the navigation system maintains full confidence in the spoofed signal. LiDAR interference can create phantom obstacles or blind perception systems. IMU drift from electromagnetic interference corrupts orientation data gradually enough to evade threshold-based detection. SATA addresses these vulnerabilities through formal, evidence-theoretic trust evaluation.

Dempster-Shafer Trust Computation

SATA uses Dempster-Shafer evidence theory rather than probability theory because it explicitly represents uncertainty through the mass assigned to the full frame of discernment Θ. This means the system can distinguish between "I trust this sensor" (high m({Trusted})), "I distrust this sensor" (high m({Compromised})), and "I do not have enough evidence" (high m(Θ)).

Trust Diagnostic Components

Asymmetric Trust Dynamics

Trust decay is deliberately faster than trust recovery: decay occurs with a 0.5-second time constant while recovery requires a 5.0-second time constant. This asymmetry ensures that a compromised sensor cannot quickly regain trust simply by momentarily producing correct readings. The 10:1 recovery-to-decay ratio means the system is conservative about trusting sensors that have been flagged as potentially unreliable.

SATA Trust Simulation

The SATA simulation provides real-time visualization of the complete trust evaluation pipeline. Users can manipulate individual sensor health, inject specific fault types, and observe how trust propagates through the Dempster-Shafer combination to produce the fused trust scalar that drives HMAA authority decisions.

Hardware Platform Integration

API Implementation

POST /trust/evaluate

{
  "sensors": [
    {"id": "lidar", "internal": 0.95,
     "cross": 0.88, "temporal": 0.92,
     "physical": 0.97, "weight": 0.30},
    {"id": "camera", "internal": 0.42,
     "cross": 0.35, "temporal": 0.50,
     "physical": 0.88, "weight": 0.25},
    {"id": "imu", "internal": 0.98,
     "cross": 0.91, "temporal": 0.96,
     "physical": 0.99, "weight": 0.20}
  ]
}

{
  "fused_trust": 0.62,
  "per_sensor": {
    "lidar": {"trust": 0.93, "status": "healthy"},
    "camera": {"trust": 0.38, "status": "degraded"},
    "imu": {"trust": 0.96, "status": "healthy"}
  },
  "disagreements": [
    {"pair": "lidar-camera", "delta": 0.55}
  ],
  "veto_active": true,
  "veto_source": "camera"
}

Minimal SATA Core

def sata_fuse(sensors):
    bpas = [build_bpa(s) for s in sensors]     # Per-sensor BPA
    cross_validate(sensors)                     # Disagreement penalty
    fused = dempster_combine(bpas)              # DS combination
    tau = fused.belief({Trusted})                # Extract trust scalar
    return clamp(tau, 0.0, 1.0)                 # τ ∈ [0,1] → HMAA

Selected References

• Shafer, G. (1976). A Mathematical Theory of Evidence. Princeton University Press.
• Khaleghi, B. et al. (2013). Multisensor Data Fusion: A Review of the State of the Art. Information Fusion, 14(1), 28-44.
• NIST (2023). AI Risk Management Framework (AI RMF 1.0).
• Cho, J.-H. et al. (2011). A Survey on Trust Management for Mobile Ad Hoc Networks. IEEE Comms Surveys, 13(4).
• Parkinson, B.W. & Enge, P. (1996). Differential GPS. In Global Positioning System, Vol. II. AIAA.

Provable Guarantees

Known Limitations and Failure Modes

Simulation Reproducibility

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

Cite This Work

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

@misc{oktenli2026sata,
  author       = {Oktenli, Burak},
  title        = {Sensor Attestation and Trust Anchoring},
  year         = {2026},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.18936251},
  url          = {https://doi.org/10.5281/zenodo.18936251},
  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). It is published on Zenodo with DOI 10.5281/zenodo.18936251 under CC BY 4.0.

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