Unveiling Client Privacy Leakage from Public Dataset Usage in Federated Distillation

The source code of the paper "Unveiling Client Privacy Leakage from Public Dataset Usage in Federated Distillation", in Proceedings on Privacy Enhancing Technologies(PoPETs 2025).

Usage

Experimental procedure:

• STEP 1: Perform the Federated Distillation
• STEP 2: Perform the Label Distribution Attack
• STEP 3: Perform the Co-op LiRA/Distillation-based LiRA.

Configure the environment needed for the experiment

pip install -r requirements.txt

• Create folders to store the MIA results

cd Privacy_PDAFD
mkdir result

Example(FedMD, CIFAR10(Private/Public), alpha = 1)

• STEP 1: Perform the Federated Distillation

python main.py --dataset 'CIFAR10' --batch_size 64  --digest_batch_size 128 --num_rounds 10 --lr 0.001 --co_lr 0.001 --num_clients 10 --num_classes 10 --sampling_rate 1 --pre_ep 20 --digest_ep 10 --revisit_ep 3 --beta 1 --seed 42 --alg 'FedMD' --model 'resnet18'  --public_part 0.2

• STEP 2: Perform the Label Distribution Attack

python ldia.py

• STEP 3(Co-op LiRA): Perform the Co-op LiRA.

python mia.py --alg FedMD --beta 1 --dataset CIFAR10 --epoch 0 --num_shadows 16 --attack coop

• STEP 3(Distillation-based LiRA): Perform Distillation-based LiRA.

python mia.py --alg FedMD --beta 1 --dataset CIFAR10 --epoch 0 --num_shadows 16 --attack distill

Details

• STEP 1: Train the target model

python main.py --dataset [dataset] --batch_size [batch_size] --digest_batch_size [batch_size] --num_rounds [rounds] --lr [learning_rate] --co_lr [learning_rate]
               --num_clients [clients] --num_classes [classes] --sampling_rate [sampling rate] --pre_ep [epochs] --digest_ep [epochs] --revisit_ep [epochs] 
               --beta [alpha_value] --seed [random seed] --alg [FD_alg] --model [model_architecture]  --public_part [public_dataset_ratio]


optional arguments:
    --dataset              //experiment dataset
    --batch_size           //batch size for local updates phase
    --digest_batch_size    //batch size for knowledge distillation phase
    --num_rounds           //number of collaboration training rounds
    --lr                   //learning rate for local updates phase
    --co_lr                //learning rate for knowledge distillation phase
    --num_clients          //number of clients
    --num_classes          //number of classes for training dataset
    --sampling_rate        //the ratio of activated clients
    --pre_ep               //training epochs for local updates phase(round 0)
    --digest_ep            //training epochs for knowledge distillation phase
    --revisit_ep           //training epochs for local updates phase(after round 0)
    --beta                 //alpha value for Dirichlet distribution
    --seed                 //random seed
    --alg                  //Federated Distillation Algorithm
    --model                //Client's model architecture
    --public_part          //public dataset size(dataset size*public_part)

• STEP 2: Perform the Label Distribution Attack

python ldia.py

• STEP 3: Perform the Co-op LiRA/Distillation-based LiRA.

python mia.py --alg [FD alg] --beta [alpha value] --dataset [dataset] 
              --epoch [target collaborated training round] --num_shadows [number of shadow models] --attck [attack method]

optional arguments:
  --alg                  //Federated Distillation Algorithm
  --beta                 //alpha value for Dirichlet distribution
  --epoch                //the attack epoch(collaborated training round)
  --num_shadows          //number of shadow models
  --attack               //attack methods(coop, distill)