Table 2 Experimental environment and hyperparameter settings.
From: Secure multi-party test case data generation through generative adversarial networks
Parameter / Component | Specification / Value |
|---|---|
Hardware Configuration | |
Server (Coordinator) | NVIDIA A100 GPU (80GB), Intel Xeon Gold 6248R |
Client (Participant) | NVIDIA Jetson AGX Xavier (32GB ARM64) \(\times\) 5 |
Network Simulation | Linux TC (Delay: 10-50ms, Bandwidth: 50-200Mbps) |
Model Architectures | |
Autoencoder (AE) | Encoder: FC(512, ReLU) \(\rightarrow\) BiLSTM(256) \(\rightarrow\) FC(32, Sigmoid) |
Decoder: FC(256, ReLU) \(\rightarrow\) LSTM(512) \(\rightarrow\) FC(d, Linear) | |
Generator (G) | Input(\(z \in \mathbb {R}^{32}\)) \(\rightarrow\) FC(128, LeakyReLU) \(\rightarrow\) LSTM(256) \(\rightarrow\) FC(d, Tanh) |
Discriminator (D) | Input(\(x \in \mathbb {R}^{d}\)) \(\rightarrow\) FC(256, LeakyReLU) \(\rightarrow\) FC(128) \(\rightarrow\) Output(1, Sigmoid) |
Software Stack | |
Frameworks | TensorFlow Federated 0.19, PySyft 0.5, OpenSSL 3.0 |
Encryption Library | Python-Paillier (Key size: \(N=2048\) bits) |
Protocol Simulators | ModbusPal, PyModbus, Eclipse Milo (OPC UA) |
Training Settings | |
Federated Learning | Global Rounds: \(T=100\), Local Epochs: \(E=5\) |
GAN Learning Rate | Generator: \(2 \times 10^{-4}\), Discriminator: \(2 \times 10^{-4}\) |
Optimizer | Adam (\(\beta _1=0.5, \beta _2=0.999\)) |
Batch Size | Local Batch: \(B=64\), Test Batch: 1,000 |
DP Parameters | Privacy Budget \(\epsilon =0.5\), \(\delta =10^{-5}\), Clipping Norm \(C=1.0\) |
Loss Weights | Adversarial: \(\lambda _{adv}=1.0\), Syntax: \(\lambda _{syn}=0.3\), AE \(\lambda _{rec}=1.0\) |
