The detector converts delayed utilization and delay measurements into bounded normalized anomaly scores. These quantities are denoted as anomaly scores rather than statistical z-scores because they are referenced to fixed operating thresholds rather than estimated sample means and standard deviations. For delayed measured utilization ( u ˜ e ) and delayed measured latency ( d ˜ e ), the two anomaly components are

where ( ℓ e ) is the nominal propagation latency of link (e). Detector evidence is

The utilization baseline of 0.62 represents the nominal utilization level above which congestion evidence begins to accumulate. For delay, evidence begins when measured latency exceeds 1.15 times nominal propagation latency. Both signals are clipped to [0, 1].

A detector-quality-specific calibration function is applied to the anomaly evidence:

where ( c q ) is 0.95, 0.80, or 0.62 for the strong, moderate, and weak detector settings, respectively. Link risk is then updated using

where ( δ q ) is the detector delay.

To simulate imperfect detection, an attack-associated evidence value is attenuated to 12% of its original magnitude when a configured false-negative event occurs. For an unattacked link experiencing a configured false-positive event, alert-like evidence between approximately 0.58 and 0.93 is injected before calibration. These stochastic error events use the probabilities reported in Table 2.

For reporting detector precision and recall, a link is classified as detector-positive when ( r e ( t ) ≥ 0.50 ). Precision and recall are computed over link-epoch decisions during attack-active evaluation epochs by comparing this detector output with evaluator-only attack ground truth. Ground-truth attack labels are used only by the simulation generator to impose controlled false-negative/false-positive events and by the evaluator to score detector performance; they are never supplied to the routing controller.

Table 2.Detector quality scenarios.

For a candidate path (p), let

denote nominal propagation latency,

denote the maximum predicted utilization after accounting for provisional traffic ( v e ) already assigned during the current routing pass, and

denote mean detector-derived path risk.

The TARA path score is

Candidate paths are initially ordered by ( J ( p ) ). The lowest-cost path becomes the primary path. When path splitting is enabled, the secondary path is selected from the remaining candidates by first minimizing the number of links shared with the primary path and then minimizing ( J ( p ) ) among equal-overlap candidates. The no-overlap ablation omits this shared-link preference. Traffic shares across the selected paths are proportional to ( 1 / J ( p ) ) and normalized to sum to one.

Table 3 identifies what each policy does with the common information stream. Static and topology-disjoint routing intentionally ignore telemetry after initialization; this is a policy restriction, not unequal information delivery. Reactive routing uses the same risk signal as TARA and changes a route only when its current path contains a link whose risk is at least 0.55.

Table 3. Policies and ablations.

For each seed, 36 nodes are placed independently and uniformly in a 1000 × 1000 two-dimensional coordinate region. Each node is connected to its four nearest neighbors, after which 18 additional undirected node pairs are selected uniformly without replacement to increase path diversity. Duplicate undirected links are merged. Propagation latency for a link between nodes u and v is generated as

1 + d { u v } 190 + ξ ms, where d u v is Euclidean distance and ξ ∼ U ( 0 , 1 ) . Capacities of

nearest-neighbor links are sampled from U ( 125 , 235 ) , while the additional random links use U ( 115 , 220 ) .

Each topology contains 90 legitimate demands. Source and destination nodes are sampled uniformly from distinct node pairs. The offered rate for each demand is sampled from a lognormal distribution with log-location 1.05 and log-scale 0.42 and is clipped to the interval 1 - 8 simulation rate units.

Initial shortest paths are computed using propagation latency as link cost. For every demand, up to four candidate paths are generated through repeated shortest-path searches in which links used by an already discovered path have their temporary cost multiplied by 2.0. This produces alternative paths while retaining a bounded candidate set.

Attack targets are selected from links that occur most frequently in the initial shortest paths. Specifically, links are ranked by the number of baseline paths that

traverse them, and the highest-ranked max ( 5 , | E | 15 ) links are selected as attack

targets. The target set remains fixed within a seed-scenario trial.

The simulation contains 10 routing epochs, with attacks active during epochs 4 - 10. For every attacked link (e) and attack-active epoch (t), attack load is generated as

where ( C e ) is link capacity, α ∈ { 0.60 , 1.00 , 1.40 } is the configured attack-intensity factor, and F e ( t ) ∼ U ( 0.82 , 1.18 ) introduces epoch-level fluctuation. Random streams are deterministically derived from the simulation seed, attack intensity, and detector setting so that policies evaluated within a scenario experience paired underlying conditions.

Table 4 gives the complete design. Each seed fixes node coordinates, capacities, demands, attack targets, measurement noise, false-alarm draws, and attack fluctuations. Policies within that scenario are therefore paired.

Table 4.Executed simulation design.

The 95% paired interval for a comparison uses Equation (5), where each difference is computed between policies sharing a seed and scenario.

The two-sided 95% paired confidence interval is computed using the Student-(t) distribution:

where Δ ¯ is the mean paired difference, s Δ is the sample standard deviation of the paired differences, and n = 24 . For the reported comparisons, d f = 23 and t { 0.975 , 23 } = 2.069 .

The interval is descriptive for the executed seed population; no claim is made that synthetic topology seeds are a random sample of deployed networks.

Attack traffic and legitimate routed traffic contribute jointly to link load. For each link e , utilization is

For a legitimate path (p), the deliverable fraction is limited by its most capacity-constrained link. The delivered fraction associated with each path share is therefore computed from

Demand delivery is the weighted sum of the fractions delivered through its selected paths.

The performance-latency model uses total utilization to approximate congestion-induced delay. For link e , the queueing factor is

The primary results aggregate only the seven attack-active epochs (epochs 4 -10). Trial-level delivered demand is the mean delivery percentage across those epochs. At each epoch, the 95th percentile of modeled path latency is computed across the 90 legitimate demands; the reported trial-level P95 latency is the 95th percentile of those seven epoch-level P95 values.

Dropped or partially undelivered demand is represented through the delivery-ratio metric and is not assigned an artificial infinite or penalty latency. The latency statistic therefore represents modeled path delay under the offered load rather than packet-completion latency for successfully delivered traffic only.

Figure 3 and Table 5 report the moderate-detector, 100%-attack scenario. Full TARA delivered 97.25% (SD 1.20) and reactive rerouting delivered 93.18% (SD 2.52). The paired difference was 4.07 percentage points, with a two-sided 95% paired t-interval from 3.39 to 4.74 percentage points.

Figure 3. Mean delivered demand for the principal scenario across 24 paired seeds.

Table 5. Principal scenario results.

Full TARA’s mean 95th-percentile latency was 27.15 ms. Its paired difference from reactive routing was −6.47 ms (two-sided 95% paired t-interval: −7.94 to −5.00 ms); negative values favor TARA.

Figure 4 shows degradation as detector quality weakens. Full TARA delivery was 97.89% with strong detection and 96.71% with weak detection. Table 6 pairs those service results with the detector precision and recall produced by the simulated measurement pipeline.

Figure 4.Delivery under strong, moderate, and weak detector settings. Axis values 1, 2, and 3 denote strong, moderate, and weak.

Table 6.Full TARA detector sensitivity at 100% attack.

Figure 5 reports all attack levels under moderate detection. The separation among policies grows as capacity pressure increases, but detector delay means that no telemetry-driven policy avoids the initial attacked epochs.

Figure 6 compares Full TARA with its four structural ablations in the principal scenario. The single-path ablation has the largest latency, showing that diversity rather than the risk term alone explains much of the resilience.

Figure 5. Delivered demand under 60%, 100%, and 140% target-capacity attack load with moderate detector quality.

Figure 6. Mean 95th-percentile latency for full TARA and four ablations in the principal scenario.

Figure 7 confirms that strong detection improves both precision and recall; weak detection reduces recall and increases false-alert exposure. Figure 8 shows that this additional adaptivity requires substantially more route changes than reactive rerouting.

Figure 7.Mean detector precision and recall obtained from the simulated telemetry pipeline.

Figure 8.Mean controller route changes in the principal scenario.