Technical Summary: Bayesian Trust and Action Fusion (BTAF) for Risk-Adaptive Access Control

1. Problem Statement

Static access control policies are often ill-suited for Internet of Things (IoT) systems where conditions fluctuate significantly between sessions. In managed IoT environments, a single access request may involve heterogeneous evidence, including biometric confidence, session token integrity, device security posture, channel health, resource stress, and the sensitivity of the requested action. Treating all sessions identically creates a trade-off between excessive user friction (over-restriction) and weak protection (under-restriction).

While risk-based authentication and zero-trust principles advocate for contextual adaptation, existing approaches often lack a lightweight mechanism to fuse these diverse signals without requiring a massive end-to-end security architecture. Furthermore, in resource-constrained IoT devices, aggressive security interventions (such as hard restrictions) can disrupt critical sensing or control workflows. The core problem addressed is how to combine heterogeneous access evidence to distinguish malicious sessions from benign ones while minimizing unnecessary operational costs, specifically avoiding hard restrictions when a lower-cost "step-up" verification is sufficient.

2. Methodology: BTAF

The paper proposes Bayesian Trust and Action Fusion (BTAF), a lightweight, dual-output method designed for resource-constrained environments. The system operates on a six-dimensional feature vector $x_i = [b_i, t_i, d_i, c_i, e_i, s_i]^\top$, representing:

• Biometric confidence ($b$)
• Token/session integrity ($t$)
• Device posture ($d$)
• Channel health ($c$)
• Resource stress ($e$)
• Action sensitivity ($s$)

The methodology consists of three primary components:

2.1 Bayesian Risk Classifier

The first component utilizes a Gaussian Naive Bayes classifier to estimate the probability that a session is malicious ($p_i$). The model assumes class-conditional features follow a Gaussian distribution. This component is intentionally simple to ensure computational efficiency and transparency, avoiding high-capacity black-box architectures.
$$P(y | x) \propto P(y) \prod_{j=1}^{6} P(x_j | y)$$
where $y=1$ denotes a malicious session.

2.2 Bayesian Intervention-Cost Model

The second component employs Bayesian ridge regression to estimate the expected operational cost of a strong intervention ($\hat{o}_i$). This model accounts for the current resource and action context to predict how disruptive a hard restriction would be.
$$o_i = w^\top x_i + \epsilon_i$$
This regression head does not determine if intervention is needed, but rather how severe the intervention should be.

2.3 Calibrated Action Selection Policy

The system maps the risk probability ($p_i$) and estimated cost ($\hat{o}_i$) to one of three actions: allow, step-up verification, or restrict.

1. Thresholding: A security threshold $\tau_f$ is established using the 90th percentile of risk scores from benign validation sessions. If $p_i < \tau_f$, the session is allowed.
2. Cost-Aware Escalation: If $p_i \ge \tau_f$, the system evaluates the estimated cost $\hat{o}_i$. A dynamic restriction threshold $\tau_r(\hat{o}_i)$ is calculated:
   $$\tau_r(\hat{o}_i) = \text{clip}(0.76 + 0.16\hat{o}_i, 0.76, 0.92)$$
3. Decision:
   • If $p_i \ge \tau_r(\hat{o}_i)$, the action is restrict (hard denial).
   • Otherwise, the action is step-up (intermediate challenge).

This structure ensures that the regression head only modulates the severity of the response for sessions already flagged as risky, preventing the cost model from lowering the initial security bar.

3. Experimental Setup and Evaluation

The study utilizes a reproducible synthetic evaluation rather than real-world deployment data.

• Data Generation: A simulator generates six scenario classes: benign operation, credential theft, token replay, device compromise, channel anomaly, and mixed low-and-slow manipulation.
• Regimes:
  • Development Regime: 15,000 sessions per seed (2,500 per scenario) representing stable operation.
  • Stress Regime: 6,000 sessions per seed representing shifted conditions (reduced biometric/device/channel scores, increased resource pressure) to simulate sensor noise, aging devices, or environmental changes.
• Protocol: Experiments were repeated across five independent seeds. Data was split into 60% training, 20% validation, and 20% test.
• Baselines: BTAF was compared against a transparent static weighted-risk score, logistic regression, and Gaussian Naive Bayes without cost-aware action selection.

4. Key Results

4.1 Development Regime Performance

In stable conditions, the BTAF risk model achieved:

• AUROC: $0.971 \pm 0.002$
• Balanced Accuracy: $0.908 \pm 0.005$
• Attack Response Rate: $0.924 \pm 0.011$
• Benign Friction: $0.107 \pm 0.019$

These results indicate that heterogeneous evidence provides useful separation and that a lightweight Bayesian model is sufficient to exploit it in the synthetic setting.

4.2 Impact of Cost-Aware Fusion

The primary benefit of the cost-aware layer was observed in the reduction of unnecessary hard restrictions:

• Benign Hard Restriction Rate: Reduced from 0.107 (binary policy) to 0.052.
• Attack Response: Remained constant at 0.924.
  This demonstrates that the system successfully distinguishes between sessions requiring additional evidence (step-up) and those requiring immediate denial, effectively halving the friction for benign users without compromising attack detection.

4.3 Distribution Shift (Stress Regime)

When operating conditions changed without recalibration:

• Attack Response: Remained high ($0.952 \pm 0.004$).
• Benign Friction: Increased significantly to 0.473 $\pm$ 0.029.
• Balanced Accuracy: Dropped to $0.740 \pm 0.013$.

This highlights a critical limitation: the method is not invariant to operating regime changes. Significant distribution shifts cause a sharp rise in false positives (benign friction), necessitating periodic recalibration.

5. Claims and Significance

The paper makes a deliberately narrow and modest claim. It does not propose a replacement for biometric systems, zero-trust architectures, device attestation, or quantum-secure communication. Instead, the significance of BTAF lies in:

1. Transparent Fusion: It provides a reproducible demonstration of how to combine heterogeneous evidence (identity, device, channel, resource, action) using a transparent probabilistic framework.
2. Separation of Concerns: It successfully decouples risk estimation (is this session malicious?) from response severity (how disruptive should the countermeasure be?).
3. Operational Efficiency: It shows that incorporating an operational cost estimate can significantly reduce unnecessary hard restrictions on benign sessions while maintaining high attack response rates.
4. Realistic Limitations: The study explicitly acknowledges that the method requires recalibration when operating conditions change materially. It treats the rise in benign friction under distribution shift not as a minor bug, but as a core deployment risk that requires governed security operations (e.g., periodic threshold validation) rather than automatic optimization.

In conclusion, BTAF serves as a lightweight, interpretable mechanism for managed IoT access, demonstrating that probabilistic risk and operational-cost estimates can be effectively separated and combined to create a three-level access policy that balances security with usability in resource-constrained environments.