Running AuditML on Google Colab¶

Google Colab provides free T4/A100 GPUs that are useful for CIFAR-10/100 experiments.

Setup¶

Open a new Colab notebook and run:

# Install AuditML and dependencies
!pip install auditml opacus torchvision maturin

# Clone the repo (to get configs and scripts)
!git clone https://github.com/EemanAsghar/AuditML-Privacy-Toolkit.git
%cd AuditML-Privacy-Toolkit

Check GPU availability¶

import torch
print(torch.cuda.is_available())          # True on Colab GPU
print(torch.cuda.get_device_name(0))      # Tesla T4

CIFAR-10 experiment¶

# Run with the CIFAR-10 config
!auditml train --config configs/audit_cifar10.yaml
!auditml audit --config configs/audit_cifar10.yaml --attack mia_threshold

Or use the full experiment script:

!python scripts/run_all_attacks.py --full

This runs all 3 datasets × 4 privacy levels × 4 attacks and writes a master CSV.

Mounting Google Drive (optional)¶

Save results to your Drive so they persist after the session ends:

from google.colab import drive
drive.mount('/content/drive')

# Update output path in config
import yaml

with open("configs/audit_cifar10.yaml") as f:
    cfg = yaml.safe_load(f)

cfg["output"]["results_dir"] = "/content/drive/MyDrive/AuditML/results"

with open("configs/audit_cifar10_colab.yaml", "w") as f:
    yaml.dump(cfg, f)

Transfer results back to local machine¶

from google.colab import files

# Zip results
!zip -r results.zip results/

# Download
files.download("results.zip")

Colab-specific tips¶

Use num_workers: 0 in your YAML config on Colab (DataLoader workers can cause issues).
Colab sessions disconnect after ~90 minutes of inactivity — use Drive to checkpoint.
For long runs, use Colab Pro or run scripts/run_all_attacks.py in quick mode first.

Quick mode vs full mode¶

# Quick mode (default) — 1 epoch, fast sanity check
!python scripts/run_all_attacks.py

# Full mode — production-quality results
!python scripts/run_all_attacks.py --full

Quick mode is useful to verify everything works before committing to a long full run.