README.md

fyp_adv_collusion_tracing

This is the repo for our final year project on " Identification of Collusive Adversaries from a Single Adversarial Example via Machine Learning Watermarks", supervised by Professor Cheng Minhao.

Baseline in paper "Identification of the Adversary from a Single Adversarial Example"

To conduct baseline experiments, cd to src/adv_tracing and follow the instructions in the README file.

To evaluate the transferability of adv attacks:

python adv_evl.py --model_name ResNet18 --dataset_name CIFAR10 --attack_name SimBA-px

Collusive Adv Attack (Our FYP project)

Firstly, cd to src/collusion_tracing.
To train a base model, run:

python train_base_model.py --model_name ResNet18 --dataset_name CIFAR10

To train different model copies, run: (You may change: model_name, dataset_name, head_id, we need 10 heads for each setting)

python train_new_and_acc_code_one_head.py --model_name ResNet18 --dataset_name CIFAR10 --num_epochs 100 --batch_size 1024 --head_id 0

To generate collusive adversarial samples, run: (You may change model_name and dataset_name)

python -m collusion_attacks.nes --model_name ResNet18 --dataset_name CIFAR10 -k 2 --num_samples 500 -b 64 --num_models 10

To evaluate the transferability of the collusive attacks, run:

python collusion_evl.py --model_name ResNet18 --dataset_name CIFAR10 --attack_name NES --collusion_attack mean

To calculate the collusive tracing accuracy， run tracing_new.ipynb You may modify the arguments in class Agrs with your {model_name, dataset_name, collusion_attack, num_obtained_adv, attack_name}