IDS for WiFi — Federated Deep Belief Network (DBN)

An intrusion detection system (IDS) for WiFi networks that combines federated learning with a Deep Belief Network (DBN). Multiple clients train locally and share model updates with a central server, which aggregates them using Federated Averaging (FedAvg).

Key features

• Federated learning: clients train locally; only model updates are shared
• DBN-based model: RBM-stacked pretraining with FNN classifier
• RPC server/client: lightweight Go services to exchange model parameters
• Visualization & metrics: accuracy, precision/recall, F1, ROC, confusion matrix

---

Repository layout

config/                      # YAML configs (placeholders)
datasets/                    # Data and processed CSVs
logs/                        # Server and client logs
output_images/               # Plots generated during training
scripts/                     # Data helpers (placeholders)
src/
  clients/
    client1/                 # Go RPC client + Python runner
    client2/, client3/       # Python clients (placeholders)
  FEDavg/                    # FedAvg aggregator + client config JSON
  models/                    # DBN/RBM implementations and training scripts
  server/                    # Go RPC server + helper scripts
  trained models/            # Local/global model checkpoints (.pth)

Notable files:

• src/server/server.go: RPC server (listens on 0.0.0.0:8080) serving and accepting model files
• src/clients/client1/client.go: RPC client to download global model and run DBN_client.py, then upload updates
• src/FEDavg/fedavg.py: FedAvg aggregator working from client_configuration.json
• src/models/{DBN.py,RBM.py}: model code; DBN_client*.py training entry points
• src/FEDavg/client_configuration.json: source of truth for clients and global model path

---

Requirements

• Python 3.10+ (3.12 tested by bytecode present)
• Go 1.20+
• Recommended: Linux environment with virtualenv/conda

Python packages commonly used by the codebase:

• torch, torchvision (optional), tqdm, pandas, scikit-learn, matplotlib, seaborn, networkx, torchviz

If requirements.txt is empty, install manually:

python -m venv .venv && source .venv/bin/activate
pip install torch tqdm pandas scikit-learn matplotlib seaborn networkx torchviz

---

Data

• Place your processed CSVs under datasets/processed/ (e.g., Filtered_dataset.csv, Preprocessed_set*.csv).
• Some scripts contain absolute paths. Update them to point to files under datasets/processed/ on your machine.
• Example quick check helper: datasets/processed/preprocess.py prints unique label values from the last column.

---

Environment variables

Set these before running server/clients to avoid path issues (note: folder name has a space):

# Run from repository root
export PARAMS_PATH="$PWD/src/trained models/GLOBAL_dbn_rbm_model.pth"
export JSON_PATH="$PWD/src/FEDavg/client_configuration.json"
export UNIQUE_ID="client_001"
export NUMBER_ITR="10000"      # number of instances trained last iteration
export NAME="$(hostname)"      # client display name

Helper scripts:

• src/server/server_setup.sh exports PARAMS_PATH and JSON_PATH to ~/.bashrc
• src/clients/client_setup.sh shows an example client-side setup and cron scheduling

---

Running the RPC server (Go)

The server expects to find auth.json and writes into ../FEDavg/ relative to its working directory.

# From repo root
cd src/server
go run server.go
# The server listens on 0.0.0.0:8080 and serves model at $PARAMS_PATH

Authentication list: src/server/auth.json (valid client IDs). Update it to match your clients.

---

Running a client (Go + Python)

The Go client performs two operations:

• dwn: download the global model (via RPC) into PARAMS_PATH and run ../../models/DBN_client.py
• updt: upload the locally updated model in PARAMS_PATH back to the server

# From repo root
cd src/clients/client1
# Set env vars as shown above, ensure dataset paths in Python scripts are valid
# 1) Download global params and trigger Python training
go run client.go -opn dwn
# 2) After training finishes, upload updated params
go run client.go -opn updt

RPC address: The client currently dials a placeholder host. Update both occurrences in src/clients/client1/client.go to your server:

• rpc.Dial("tcp", "NotAvailable-37552.portmap.host:37552") → e.g. rpc.Dial("tcp", "<server-ip>:8080")

---

Federated averaging (FedAvg)

Aggregation is driven by src/FEDavg/client_configuration.json:

• network_summary.last_iteration_summary.global_model_path: path to the global model checkpoint
• clients[]: list of clients with instances_trained_last_iteration and their model checkpoints

You can trigger FedAvg manually:

# Ensure JSON_PATH is exported to src/FEDavg/client_configuration.json
python src/FEDavg/fedavg.py

Note: The server also invokes FedAvg after receiving a client update.

---

Centralized or standalone training

Several training entry points are provided under src/models/. Make sure to change any absolute dataset paths to your local paths under datasets/processed/.

• DBN_client.py, DBN_client_with_visualization.py, DBN_client_two_attacks.py
• DBN_server.py (server-side training)

These scripts typically:

• Pretrain a DBN via stacked RBMs (see RBM.py)
• Build an FNN from the DBN layers (initialize_model)
• Train and report metrics, optionally saving a .pth checkpoint to src/trained models/

---

Outputs

• Checkpoints: src/trained models/*.pth
• Metrics/plots: output_images/*.png
• Logs: logs/server.log, logs/client*.log

---

Troubleshooting

• RPC cannot connect: Update the RPC host/port in src/clients/client1/client.go to your server’s IP and open port 8080.
• Relative paths fail: Run the server from src/server/; it expects auth.json in CWD and writes to ../FEDavg/.
• Spaces in path: The trained models folder contains a space. Always quote PARAMS_PATH as shown.
• FedAvg script error: If src/FEDavg/fedavg.py complains about function arguments, ensure _load_json signature matches how it is called and that JSON_PATH points to an existing JSON.
• Hard-coded absolute paths: Some model scripts include absolute paths. Replace them with repo-relative paths.

---

Contributing

• Open an issue describing the change
• Create a feature branch and submit a PR
• Please avoid committing large binaries; link to dataset sources instead

License

MIT License