Algorithm 1: Dynamic Threat-Aware Quantization

Require: Current threat level $\tau \left(\mathrm{t}\right)$, thresholds ${\tau }_{\mathrm{l}\mathrm{o}\mathrm{w}},{\tau }_{\mathrm{m}\mathrm{e}\mathrm{d}},{\tau }_{\mathrm{h}\mathrm{i}\mathrm{g}\mathrm{h}}$ Ensure: Selected quantization bit-width ${\mathrm{q}}_{\mathrm{b}}\left(\mathrm{t}\right)$     If τ(t) <τlow then   ${\mathrm{q}}_{\mathrm{b}}\left(\mathrm{t}\right)\leftarrow 4$   Else if τ low ≤ τ(t) < τmed then   ${\mathrm{q}}_{\mathrm{b}}\left(\mathrm{t}\right)\leftarrow 8$   Else if τmed ≤ τ(t) < τhigh then   ${\mathrm{q}}_{\mathrm{b}}\left(\mathrm{t}\right)\leftarrow 16$ Else   ${\mathrm{q}}_{\mathrm{b}}\left(\mathrm{t}\right)\leftarrow 32$   End if   Return  ${\mathrm{q}}_{\mathrm{b}}\left(\mathrm{t}\right)$    Update parameters: $\theta \leftarrow \theta -\eta \cdot \nabla \theta \mathrm{L}$  end for Validate model on held-out validation set Apply learning rate decay if validation loss plateaus end for Return $\theta$