Algorithm 3 Local model training at fog layer

Require:   1: Set of controllers $C=\{C_1,C_2,\dots ,C_n\}$   2: Time step t, Learning rate $\mu$, Batch size B, Regularization parameter $\lambda$   3: Dataset $D_{ctrl}^t=\{d_1^t,d_2^t,\dots ,d_n^t\}$ from east–west traffic telemetry   4: Initial local malicious detection models:   5: CNN-based model ${\mathcal{T}}_{\mathrm{CNN},i}^t$, SVM-based model ${\mathcal{T}}_{\mathrm{SVM},i}^t$, General local model ${\mathcal{T}}_i^t$   6: Controller private keys ${\kappa }_{priv}^i$ and public keys ${\kappa }_{pub}^i$ Ensure:   7: Updated and signed local models ${\mathcal{T}}_i^{t+1}$ with digital signatures ${\mathsf{\Sigma }}_i$   8: Prepared model updates for secure transmission to the blockchain layer   9: for each controller node $C_i\in C$ do   10:     Step 1: Data Collection   11:     Load local traffic data $D_{ctrl}^t$ from east–west interface flows   12:     Extract feature vectors $d_k^t\in D_{ctrl}^t$   13:     Step 2: Local Model Initialization   14:     Retrieve previous local models ${\mathcal{T}}_i^t,{\mathcal{T}}_{\mathrm{CNN},i}^t,{\mathcal{T}}_{\mathrm{SVM},i}^t$   15:     Step 3: Gradient Computation for General Model   16:     Compute gradient of loss: $\nabla \mathcal{L}({\mathcal{T}}_i^t,D_{ctrl}^t)$   17:     Step 4: General Model Update   18:     Update local model weights:   19:                ${\mathcal{T}}_i^{t+1}\leftarrow {\mathcal{T}}_i^t-\mu ·\nabla \mathcal{L}({\mathcal{T}}_i^t,D_{ctrl}^t)$   20:     Step 5: CNN Model Training   21:     For each batch $j=1\dots B$ do   22:                Extract batch samples $(x_j,y_j)$   23:                Compute CNN gradient:   24:                     $\nabla {\mathcal{L}}_{\mathrm{CNN}}(f_{\theta }\left(x_j\right),y_j)$   25:     end for   26:     Update CNN model:   27:                ${\mathcal{T}}_{\mathrm{CNN},i}^{t+1}\leftarrow {\mathcal{T}}_{\mathrm{CNN},i}^t-\mu ·{\displaystyle \frac{1}{B}}{\sum }_{j=1}^B\nabla {\mathcal{L}}_{\mathrm{CNN}}(f_{\theta }\left(x_j\right),y_j)$   28:     Step 6: SVM Model Training with Hinge Loss   29:     For each sample $(x_j,y_j)\in D_{ctrl}^t$ do   30:                Compute indicator function:   31:                     $\mathbb{I}(y_j·\langle {\mathcal{T}}_{\mathrm{SVM},i}^t,x_j\rangle <1)$   32:                Calculate SVM gradient component:   33:                     $g_j=\lambda ·{\mathcal{T}}_{\mathrm{SVM},i}^t-y_j{x}_j·\mathbb{I}(\cdots )$   34:     end for   35:     Aggregate gradient over dataset and update:   36:                ${\mathcal{T}}_{\mathrm{SVM},i}^{t+1}\leftarrow {\mathcal{T}}_{\mathrm{SVM},i}^t-\mu ·{\sum }_j{g}_j$   37:     Step 7: Digital Signature Creation   38:     Sign updated model using private key:   39:                ${\mathsf{\Sigma }}_i\leftarrow \mathrm{Sign}({\mathcal{T}}_i^{t+1},{\kappa }_{priv}^i)$   40:     Step 8: Secure Transmission   41:     Transmit $({\mathcal{T}}_i^{t+1},{\mathsf{\Sigma }}_i)$ to blockchain layer   42:     Step 9: Signature Verification   43:     Verify signature with public key:   44:                $\mathrm{Validate}({\mathsf{\Sigma }}_i,{\kappa }_{pub}^i)$   45:     if validation accepted then   46:         Accept model update for global aggregation   47:     else   48:         Reject update, log security alert   49:     end if   50: end for