自定义侧信道攻击中的网格搜索超参数方法
在侧信道攻击中使用的超参搜索方法均为for循环，最近在训练AE时发现sklearn有现成的greph方法来完成网格调参，但是自带的方法自由度较低，且在判定超参好坏时并没有考虑到侧信道攻击并不主要以acc来判定模型的攻击效率，这里引入sklearn中的ParameterGrid方法来完成自定义构造。
Python: 3.8.16 Sklearn==1.0.2
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
| def monkey_search(data, params, model, callbacks, key, shuffle=True, pred_size=5000, verbose=1):
assert type(callbacks) == list
train_data, test_data = data['train_data'], data['test_data']
idx = None
for i, callback in enumerate(callbacks):
if callback.__class__.__name__ == 'OneCycleLR':
idx = i
candidatas = ParameterGrid(param_grid=params)
for index, candidata in enumerate(candidatas):
x_train, x_val, y_train, y_val = train_test_split(train_data[0], train_data[1], test_size=0.2, shuffle=True) if(shuffle == True) else train_test_split(train_data[0], train_data[1], test_size=0.2)
optimizer = candidata['optimizer']
learning_rate = candidata['learning_rate']
epoch = candidata['epoch']
batch_size = candidata['batch_size']
loss = candidata['loss']
if idx != None:
callbacks[idx].num_samples = x_train.shape[0]
callbacks[idx].batch_size = batch_size
callbacks[idx].max_lr = learning_rate
model.compile(optimizer=optimizer, loss=loss, metrics=['acc'])
model.optimizer.learning_rate.assign(learning_rate)
model.fit(x_train, y_train, validation_data=(x_val, y_val), callbacks=callbacks, epochs=epoch, batch_size=batch_size)
ranks = np.zeros((10, pred_size))
for i in range(ranks.shape[0]):
ranks[i] = mean_rank_score(model, test_data, pred_size=pred_size, key=key)
if verbose == 1:
show_result(ranks, optimizer, learning_rate, epoch, batch_size, loss)
|
该方法接收参数格式如下：
data：{"train_data": [x_train, y_train], "test_data": [x_test, y_test]}
params:{"optmizer": [string], "learning_rate:"[int],"epoch":[int],"batch_size":[int], "loss":[string] }
callbacks:{callback_function}
key: test_key int
model: model object
1
2
3
4
5
6
7
8
9
10
11
12
13
14
| def mean_rank_score(model, test_data, pred_size, key):
attack_traces, targets = test_data[0], test_data[1]
predictions = model.predict(attack_traces, verbose=0)
predictions = np.log(predictions + 1e-40)
rank_times = np.zeros(pred_size)
pred = np.zeros(256)
idx = np.random.randint(predictions.shape[0], size=pred_size)
for i, p in enumerate(idx):
for k in range(pred.shape[0]):
pred[k] += predictions[p, targets[p, k]]
ranked = np.argsort(pred)[::-1]
rank_times[i] = list(ranked).index(key)
return rank_times
|
1
2
3
4
5
6
7
| def show_result(rank_times, optmizier, learning_rate, epoch, batch_size, loss):
fig, ax = plt.subplots(figsize=(15, 7))
plt.title("optmizer:" + str(optmizier) + '_' + str(learning_rate) + '_epoch:' + str(epoch) + '_batch' + str(batch_size) + '_loss: '+ str(loss))
x = [x for x in range(0, 5000)]
ax.plot(x, np.mean(rank_times, axis=0), 'b')
ax.set(xlabel="number of traces", ylabel="mean rank")
plt.show()
|
完整代码在：https://github.com/i4mhmh/A-custom-gridSearch-method-in-SCA
