Polarization-Regularized Adversarial Pruning for Efficient Radio Frequency Fingerprint Identification on IoT Devices

View full-text article in PMC

. 2026 Mar 23;26(6):2005. doi: 10.3390/s26062005

Algorithm 1 Flow of the proposed AL-PP algorithm

Input:
Training dataset $X = {x_{1}, x_{2}, \dots, x_{n}}$ , where n denotes the number of training samples; fixed pre-trained baseline network $f_{b} (\cdot; M_{B})$ ; pruning network $f_{g} (\cdot; M_{G}, m)$ ; discriminator $D (\cdot; M_{D})$ ; sparsity factor $λ$ ; polarization parameter $μ$ ; maximum number of training epochs N.
1:
Phase 1: Sparsity training with performance recovery
2:
Load the parameters of the pre-trained baseline network $M_{B}$ .
3:
Initialize the pruning network parameters $M_{G}$ and the soft masks m, as well as the discriminator parameters $M_{D}$ .
4:
for $e p o c h = 1$ to N do
5:
Fix baseline network $f_{b} (\cdot; M_{B})$ and update discriminator $D (\cdot; M_{D})$
6:
Feed the training samples $x \in X$ into the baseline network and the pruning network to obtain the outputs $y_{b} = f_{b} (x; M_{B})$ and $y_{g} = f_{g} (x; M_{G}, m)$ .
7:
Update the discriminator parameters $M_{D}$ by maximizing the following objective using stochastic gradient descent (SGD):
$max_{M_{D}} L_{D} (M_{D}) + R_{D} (M_{D})$
8:
Fix discriminator $D (\cdot; M_{D})$ and update pruning network $f_{g} (\cdot; M_{G}, m)$
9:
Update the pruning network parameters $M_{G}$ and the soft masks m by minimizing the following objective using SGD:
$min_{M_{G}, m} L_{G} (M_{G}, m) + L_{M} (M_{G}, m) + R_{m} (m) + R_{G} (M_{G})$
10:
end for
11:
Phase 2: Structured channel pruning
12:
Evaluate the importance of each channel according to the learned soft mask values.
13:
Retain channels with large mask values and prune channels with small mask values to obtain the compact pruned model.
Output:
The compact model after structured pruning.