Federated learning filtering non-influential data for non-IID data

Federated learning with influence function

Federated learning with ICML 2017 Best paper "influence function" This code is compatilble with Tensorflow 2.x which can use eagar execution (from previous TensorFlow 1.x)

1 versus 7 classification In MNIST dataset problem

---

below pictures are the most harmful data for federated learning

After filtering non-influential data

---

How to calculate influence on loss ?

Step 0. Build model

model = tf.keras.models.Sequential()
model.add(tf.keras.layers.Input(shape=(784,)))
model.add(tf.keras.layers.Dense(1, activation='sigmoid', use_bias=True))
model.compile(loss='binary_crossentropy', metrics=['accuracy'])

model.fit(x, y)

Step 1. you should calculate s_test (inversed hessian vector product)

TEST_INDEX = 5

x_test_tf = tf.convert_to_tensor(x_test[TEST_INDEX: TEST_INDEX+1])
y_test_tf = tf.convert_to_tensor(y_test[TEST_INDEX: TEST_INDEX+1])

test_grad_my = grad_z(x_test_tf, y_test_tf, f=model)

Step 2. you also should calculate gradient of specific train data (grad z of train data i)

s_test_my = get_inv_hessian_vector_product(x_train_tf, y_train_tf, test_grad_my, model,
                                            scale=10,
                                            n_recursion=1000,
                                            verbose=False)
                                            

Step 3. you should multipliy gradient of train with s_test (for each train data)

for i in range(train_sample_num):
    
    # Get train grad
    train_grad = grad_z(x_train_tf[i: i+1], y_train[i: i+1], model, for_train=True)
    loss_diff_approx[i] = multiply_for_influe(train_grad, s_test_my) / train_sample_num
    

---

Estimation of loss in model without bias

Estimation of loss in model with bias

-------------------------------------------------

Calculation

---

Requirement

tensorflow 2.x

numpy 1.xx (we recommand to use 1.18.5 which we used in our paper)

sklearn 0.23.x (only used in simulation)

Install

This code was inspired by https://github.com/nayopu/influence_function_with_lissa

and ICML 2017 best paper Understanding Black-box Predictions via Influence Functions (https://arxiv.org/pdf/1703.04730)