Algorithm 1 STAG-GuardNet Training and Prediction Framework

Require:  Training set ${\mathcal{D}}_{\mathrm{tr}}$, validation set ${\mathcal{D}}_{\mathrm{val}}$, test set ${\mathcal{D}}_{\mathrm{te}}$, HHMAO population size M, iteration limit I, and candidate hyperparameter vector $\mathit{\theta }$ Ensure:  Optimized configuration ${\mathit{\theta }}^{*}$ and predicted attack-state probabilities   1:Initialize HHMAO population $\mathcal{P}={\left\{{\mathit{\theta }}^{\left(m\right)}\right\}}_{m=1}^M$   2:for$i=1$ to I do   3:   for $m=1$ to M do   4:     Configure STAG-GuardNet using candidate ${\mathit{\theta }}^{\left(m\right)}$ and initialize trainable network weights   5:     for each training epoch do   6:        for each mini-batch $\mathcal{B}\subset {\mathcal{D}}_{\mathrm{tr}}$ do   7:          Construct the telemetry-adaptive graph and compute graph aggregation $\mathbf{P}\left(t\right)$   8:          Update attack memory $\mathbf{M}\left(t\right)$ and temporal representation $\mathbf{T}\left(t\right)$   9:          Apply risk-aware attention and obtain predictions $\widehat{\mathbf{y}}\left(t\right)$ 10:          Compute weighted cross-entropy loss $\mathcal{L}$ 11:          Update trainable network weights using Adam and backpropagation 12:        end for 13:     end for 14:     Evaluate ${\mathit{\theta }}^{\left(m\right)}$ on ${\mathcal{D}}_{\mathrm{val}}$ using the validation objective $\mathcal{F}\left({\mathit{\theta }}^{\left(m\right)}\right)$ 15:   end for 16:   Update candidate configurations using HHO exploration and MRFO refinement 17:end for 18:Select ${\mathit{\theta }}^{*}$ with the best validation objective and fix the corresponding STAG-GuardNet configuration 19:Perform the forward pass on ${\mathcal{D}}_{\mathrm{te}}$ using the fixed trained model without HHMAO updates 20:return ${\mathit{\theta }}^{*}$ and test-set attack-state probabilities