Configuration¶

Simulations are controlled entirely by a single JSON file. The file has two top-level keys:

{
  "shared_settings": { ... },
  "simulation_strategies": [
    { ... },
    { ... }
  ]
}

Example¶

{
  "shared_settings": {
    "aggregation_strategy_keyword": "pid_standardized",
    "dataset_keyword": "femnist_iid",
    "num_of_rounds": 10,
    "num_of_clients": 5,
    "num_of_malicious_clients": 1,
    "training_device": "cpu",
    "cpus_per_client": 1,
    "gpus_per_client": 0.0,
    "batch_size": 20,
    "num_of_client_epochs": 1,
    "save_plots": true,
    "save_csv": true,
    "Kp": 1, "Ki": 0.05, "Kd": 0.05
  },
  "simulation_strategies": [
    { "num_std_dev": 2.33 },
    { "num_std_dev": 3.0 }
  ]
}

This runs two strategies back-to-back, differing only in num_std_dev. See config/simulation_strategies/example_strategy_config.json for a full example with all 11 attack types.

Research Integrity & Validation¶

InteFL enforces a Scientific Integrity First policy. To ensure experimental transparency and reproducibility (aligned with IEEE Std and NeurIPS/ICML checklists), the framework uses a "fail-fast" validation approach: it rejects incompatible or mathematically unsound configurations rather than attempting to silently auto-correct them.

Key Constraints¶

• Participation vs. Removal: You cannot set remove_clients: true if your configuration requires 100% participation (e.g., min_fit_clients == num_of_clients). It is mathematically impossible to maintain full participation if clients are being permanently excluded.
• Byzantine Tolerance Bounds: robust aggregation algorithms have strict breakdown points.
  • Krum / Multi-Krum: Requires \(n > 2f + 2\) (where \(n\) is total clients and \(f\) is malicious clients).
  • Trimmed Mean: Requires trim_ratio < 0.5.
  • Violating these bounds removes the theoretical guarantees of the defense and leads to undefined behavior.

Field reference¶

Core simulation¶

Training¶

Client selection¶

Attack schedule¶

See the Attacks page for full documentation of all 11 attack types (data and model poisoning).

Client removal¶

Output¶

Strategy-specific parameters¶

Trust-based¶

PID-based¶

Krum / Multi-Krum / Bulyan¶

Trimmed Mean¶

General¶

HuggingFace / custom text datasets¶

LLM / transformer options¶