Quantified Accuracy and Miss Rates for AI-Based Traffic Detection Systems

confidence: high · 2 sources

What the evidence says

The claims collected here measure detection accuracy for AI models classifying network traffic (DSCP manipulation and software-defined network intrusion detection), not visual/camera-based traffic enforcement. A single-architecture CNN classifier trained on labeled DSCP traffic (70:15:15 train/validation/test split) reaches 96.88% accuracy, 97.33% precision, 96.88% recall, and an F1-score of 96.83% on its held-out test set. The same study tests an ensemble combining CNN, RNN, and LSTM architectures on the same manipulation-detection task and reports 99.28% accuracy, the highest of the architectures evaluated in that paper — pointing to ensembling across model types as the mechanism behind the gain over the single-CNN result, though the ensemble paper does not report precision, recall, or F1 to confirm the gain holds across all metrics.

A separate SDN-focused detection framework, evaluated across multiple benchmark datasets with a Mininet simulation, reports 97.75% average detection accuracy alongside three operational-cost figures measured in the same experiment: 15 ms detection latency, 0.25 J energy consumption per detection, and less than 5% throughput degradation. The same framework is reported to cut energy usage by 40% relative to traditional CNN- and LSTM-based approaches, again citing 0.25 J per detection as the basis for that comparison.

No claim in this set reports a false-negative or miss rate directly — recall (96.88% for the CNN classifier) is the closest proxy, and it is not present for the ensemble or SDN-framework results. Nor is there any claim addressing generalization beyond benchmark or test-set conditions to real-world deployment; the source notes on both the SDN-framework accuracy claim and the latency claim explicitly flag that whether these figures hold under real-world network conditions is unresolved.

Where the evidence splits

No contradictions in the current sources.

Thresholds and dose-response

No threshold established in the current sources.

Open empirical questions

Do any of these detection accuracies hold on real-world traffic outside benchmark datasets and test-set splits?
What are the false-positive and false-negative (miss) rates by attack or manipulation type, rather than aggregate accuracy alone?
Does the ensemble model's accuracy advantage over the single CNN hold for precision, recall, and F1, not just raw accuracy?
Were the CNN and LSTM baselines used for the energy-reduction comparison tuned to a comparable standard as the proposed framework?

Findings

CNN classifier detects DSCP traffic manipulation with high accuracy, precision, and recall

confidence — direction: highconfidence — magnitude: highconsensus: thin — single source

mechanism: supervised CNN trained on a labeled dataset of normal and manipulated DSCP traffic patterns
effect: 96.88% accuracy, 97.33% precision, 96.88% recall, 96.83% F1-score on the held-out test set
conditions / moderators: measured on a 70:15:15 train:validation:test split; not established whether this holds outside the test-set distribution

Rahman MMH (2026), Scientific reports · doi:10.1038/s41598-026-44350-6 · https://europepmc.org/articles/PMC13039465?pdf=render (retrieved 2026-08-09)

Ensembling CNN, RNN, and LSTM increases detection accuracy over the single-architecture CNN model

confidence — direction: highconfidence — magnitude: mediumconsensus: thin — single source

mechanism: ensemble learning combines three deep learning architectures (CNN, RNN, LSTM) trained on the same labeled DSCP manipulation dataset
effect: 99.28% detection accuracy, versus 96.88% for the single CNN model in the same study
conditions / moderators: precision, recall, and F1 for the ensemble are not reported, so it is not established whether the accuracy gain holds across those metrics; evaluation subset is unclear

Rahman MMH (2026), Scientific reports · doi:10.1038/s41598-026-44350-6 · https://europepmc.org/articles/PMC13039465?pdf=render (retrieved 2026-08-09)

An SDN-based detection framework achieves 97.75% average detection accuracy with 15 ms latency and under 5% throughput degradation

confidence — direction: highconfidence — magnitude: medium (low for the throughput-degradation figure, reported only as an upper bound)consensus: thin — single source

mechanism: not stated beyond "the proposed framework"; evaluated via benchmark datasets and Mininet simulation
effect: 97.75% average detection accuracy, 15 ms detection latency, less than 5% throughput degradation
conditions / moderators: not established whether accuracy generalizes to non-benchmark real-world SDN traffic, whether latency holds under real-world network conditions, or whether it holds at higher-bandwidth/higher-scale deployments; not broken down by attack type or false-positive/false-negative rate

Shameli R (2026), Frontiers in artificial intelligence · doi:10.3389/frai.2026.1701944 · https://europepmc.org/articles/PMC12929375?pdf=render (retrieved 2026-08-10)

The same SDN detection framework decreases energy consumption per detection relative to traditional CNN- and LSTM-based approaches

confidence — direction: highconfidence — magnitude: mediumconsensus: thin — single source

mechanism: not stated beyond "the proposed framework," compared experimentally against CNN and LSTM baselines
effect: 0.25 J energy consumption per detection, a 40% reduction versus traditional CNN- and LSTM-based approaches
conditions / moderators: not established whether the CNN and LSTM baselines were optimized to the same degree, or whether the comparison is fair across differing model architectures and implementation details

Shameli R (2026), Frontiers in artificial intelligence · doi:10.3389/frai.2026.1701944 · https://europepmc.org/articles/PMC12929375?pdf=render (retrieved 2026-08-10)