Fig 1.
End-to-end workflow of the sparse-selective quantization framework.
Table 1.
Detection performance comparison across datasets (Mean ± Std Dev; 95% CI for FPR).
Fig 2.
Detection accuracy (F1-score) as a function of threat prevalence across different detection methods.
The proposed method maintains consistently high accuracy (>0.90) even at low threat prevalence (0.1), whereas baseline methods show significant degradation as threat frequency decreases. The performance gap between our method and the best baseline (Full-precision LSTM) widens from 2.5% at high prevalence (0.9) to 5.0% at low prevalence (0.1), demonstrating the effectiveness of sparsity-aware feature selection for rare threat detection. Error bars indicate ±1 standard deviation.
Table 2.
Computational efficiency metrics (Mean ± Std Dev).
Fig 3.
Per-sample inference latency (in milliseconds) as the number of processed samples increases.
The proposed method maintains stable sub-millisecond latency (0.9 ms) across all sample sizes, while all baseline methods exhibit increasing latency with larger sample volumes. At 10,000 samples, our method achieves 3.2× lower latency than the best baseline (Uniform 4-bit GRU: 2.5 ms) and 5.7× lower than Commercial IDS (5.1 ms), highlighting the scalability advantages of our sparse-selective quantization approach. Error bars indicate ±1 standard deviation.
Fig 4.
Heatmap of feature importance scores based on normalized L1-norm sparsity metrics across five network traffic scenarios.
Features 0–1 (protocol-related attributes) and Feature 9 (temporal patterns) exhibit consistently high importance scores (>0.8) across all scenarios, corresponding to established attack signatures in cybersecurity literature. Features 2–8 show progressive importance degradation from Scenario 0 to Scenario 4, indicating that their relevance varies with network conditions. This sparsity-guided selection enables dynamic quantization policies that preserve discriminative power while compressing less critical features.
Table 3.
Ablation study results (Mean ± Std Dev).
Fig 5.
F1-score and latency versus sparsity threshold (Error bars: ± 1 std dev).
Fig 6.
Stability of Welford’s algorithm for and
estimation (Dashed lines: batch estimation; Solid lines: Welford’s algorithm).