Public articles linked to the same research event.
arXiv Treating Internet activity in CDR grids as a proxy for hidden signaling overload, the study trains a lightweight CNN on stylized injections, stress-tests it with diffusion-synthesized surges that preserve normal traffic structure, and finds that at the default threshold F1 collapses from about 99.94% at baseline to 0% while ROC-AUC stays at 0.886, showing ranking survives but the threshold does not; hard-sample adaptation across three random seeds raises F1 to 85.67% and ROC-AUC to 0.99996, yielding a reusable pre-deployment stress test.
Treating Internet activity in CDR grids as a proxy for hidden signaling overload, the study trains a lightweight CNN on stylized injections, stress-tests it with diffusion-synthesized surges that preserve normal traffic structure, and finds that at the default threshold F1 collapses from about 99.94% at baseline to 0% while ROC-AUC stays at 0.886, showing ranking survives but the threshold does not; hard-sample adaptation across three random seeds raises F1 to 85.67% and ROC-AUC to 0.99996, yielding a reusable pre-deployment stress test.
Treating Internet activity in CDR grids as a proxy for hidden signaling overload, the study trains a lightweight CNN on stylized injections, stress-tests it with diffusion-synthesized surges that preserve normal traffic structure, and finds that at the default threshold F1 collapses from about 99.94% at baseline to 0% while ROC-AUC stays at 0.886, showing ranking survives but the threshold does not; hard-sample adaptation across three random seeds raises F1 to 85.67% and ROC-AUC to 0.99996, yielding a reusable pre-deployment stress test.
Treating Internet activity in CDR grids as a proxy for hidden signaling overload, the study trains a lightweight CNN on stylized injections, stress-tests it with diffusion-synthesized surges that preserve normal traffic structure, and finds that at the default threshold F1 collapses from about 99.94% at baseline to 0% while ROC-AUC stays at 0.886, showing ranking survives but the threshold does not; hard-sample adaptation across three random seeds raises F1 to 85.67% and ROC-AUC to 0.99996, yielding a reusable pre-deployment stress test.