6G is the next generation for the communication systems. In recent years, machine learning algorithms have been applied widely in various fields such as health, transportation, and the autonomous car. The predictive algorithms will be used in 6G problems. With the rapid developments of deep learning techniques, it is critical to take the security concern into account to apply the algorithms. While machine learning offers significant advantages for 6G, AI models’ security is ignored. Since it has many applications in the real world, security is a vital part of the algorithms. This post has proposed a mitigation method for adversarial attacks against 6G machine learning models with adversarial learning. The main idea behind adversarial attacks against machine learning models is to produce faulty results by manipulating trained deep learning models for 6G applications.
Cite The Work If you find this implementation useful please cite it:
@misc{ 6g-sec-problems,
author = "Catak, FO. ",
title = "6G with Adversarial Machine Learning Security Problems: Millimeter Wave Beam Prediction Use-Case",
year = "2021",
url = "https://arxiv.org/abs/1905.01999",
note = "[arXiv:1905.01999 ]"
}
1. Introduction to 6G
Up to 4G, time-frequency domain technology has been explored to increase overall system capacity. The recent developments in 5g and beyond technologies support emerging applications such as smart homes, vehicular networks, augmented reality (AR), virtual reality (VR) with unprecedented rates enabled by recent advances in massive multiple-input multiple-output (MIMO), mmWave communications, network slicing, small cells, and Internet of things (IoT). These complex structures of 5G and beyond technologies can be captured by using data-driven approach machine learning (ML) algorithm. The strong learning, reasoning, intelligent recognition abilities of ML allow the network structure to train and adapt itself to support diverse demands of the systems without human intervention.
On the other hand, 6G is vulnerable for the adversarial machine learning attacks. Figure shows an attack scenario. The attacker gets an legitimate input, x, creates a noise vector with an e budget sums the input instance and the craftily designed noise to create adversarial input x_adv.

2. Adversarial Machine Learning
Adversarial machine learning is an attack technique that attempts to fool neural network models by supplying craftily manipulated input with a small difference. The obvious intention is to produce a failure in a neural network.
There are many attacking methods for deep learning models, and FGSM is the most straightforward and powerful attack type. FGSM works by utilizing the gradients of the neural network to create an adversarial example to evade the model. For an input instance x.

The overall steps are following;

3. Solution: Adversarial Training
Adversarial training is a widely recommended defense that implies generating adversarial instances using the gradient of the victim classifier, and then re-training the model with the adversarial instances and their respective labels. This technique has demonstrated to be efficient in defending models from adversarial attacks.
Let us first think a common classification problem with a training instances X of dimension d, a label space Y. We assume the classifier h has been trained to minimize a loss function l as follows:

We modified the optimization function for a 6g security problem. The following figure shows the diagram of RF beamforming codeword adversarial training.

5. Let’s get coding
I will show the our proposed approach with Python language. I import the usual standard libraries to build a Deep Learning model to predict RF beamforming codeword.
import tensorflow as tf
from scipy.io import loadmat, savemat
import numpy as np
from keras.models import Sequential
from keras.layers import Dense
from tqdm import tqdm
from keras.callbacks import ModelCheckpoint
import pickle
from sklearn.metrics import mean_squared_error
import matplotlib.pyplot as plt
import pandas as pd
In this work, I will use publicly avaliable DeepMIMO dataset to attack RF beam forming prediction model. I will use FGSM attack, and Mean Squared Error loss function for the input manipulation. My FGSM attack implementation is here.
loss_object = tf.keras.losses.MeanSquaredError()
def fgsm(model, input_instance, label, n_BS,n_beams, epsilon =0.01):
tensor_input_instance = tf.convert_to_tensor(input_instance, dtype=tf.float32)
adv_x = input_instance
for idx in range(0, n_BS*n_beams-2, n_beams):
with tf.GradientTape() as tape:
tmp_label = label[:, idx:idx + n_beams]
tape.watch(tensor_input_instance)
prediction = model(tensor_input_instance)
loss = loss_object(tmp_label, prediction)
gradient = tape.gradient(loss, tensor_input_instance)
signed_grad = tf.sign(gradient)
adv_x = adv_x + eps * signed_grad
return adv_x
The parameters in our scenario
The table shows the all parameters for RF beamforming, deep neural network, and adversarial setup.

# Reading input and output sets generated from MATLAB
# with DeepMIMO generator http://www.deepmimo.net
In_set_file=loadmat('DLCB_dataset/DLCB_input.mat')
Out_set_file=loadmat('DLCB_dataset/DLCB_output.mat')
In_set=In_set_file['DL_input']
Out_set=Out_set_file['DL_output']
# Parameter initialization
num_user_tot=In_set.shape[0]
DL_size_ratio = .8
count=0
num_tot_TX=4
num_beams=512
The original (undefended) train procedure
def train(In_train, Out_train, In_test, Out_test,
nb_epoch, batch_size,dr,
num_hidden_layers, nodes_per_layer,
loss_fn,n_BS,n_beams):
in_shp = list(In_train.shape[1:])
AP_models = []
for idx in range(0, n_BS*n_beams-2, n_beams):
idx_str = str(idx / n_beams + 1)
act_func = 'relu'
model = Sequential()
model.add(Dense(100, input_dim=in_shp[0], activation=act_func))
model.add(Dense(100, activation=act_func))
model.add(Dense(100, activation=act_func))
model.add(Dense(100, activation=act_func))
model.add(Dense(n_beams, activation=act_func))
model.compile(loss=loss_fn, optimizer='rmsprop', metrics=['mean_squared_error'])
history = model.fit(In_train,
Out_train[:, idx:idx + n_beams],
batch_size=batch_size,
epochs=nb_epoch,
verbose=0,
validation_data=(In_test, Out_test[:,idx:idx + n_beams]))
filehandler = open('history.pkl', 'wb')
pickle.dump(history.history, filehandler)
filehandler.close()
AP_models.append(model)
return AP_models
Training process (Normal Behaviour)
Here we will train the RF beamforming codeword prediction model with out any attacker.
count=count+1
DL_size=int(num_user_tot*DL_size_ratio)
np.random.seed(2016)
n_examples = DL_size
num_train = int(DL_size * 0.8)
num_test = int(num_user_tot*.2)
train_index = np.random.choice(range(0,num_user_tot), size=num_train, replace=False)
rem_index = set(range(0,num_user_tot))-set(train_index)
test_index= list(set(np.random.choice(list(rem_index), size=num_test, replace=False)))
In_train = In_set[train_index]
In_test = In_set[test_index]
Out_train = Out_set[train_index]
Out_test = Out_set[test_index]
#In_train, In_test, Out_train, Out_test = train_test_split(In_set, Out_set, test_size=0.33)
# Learning model parameters
nb_epoch = 10
batch_size = 100
dr = 0.05 # dropout rate
num_hidden_layers=4
nodes_per_layer=In_train.shape[1]
loss_fn='mean_squared_error'
eps = 2.0 * 16.0 / 255.0
# Model training
AP_models = train(In_train, Out_train, In_test, Out_test,
nb_epoch, batch_size,dr,
num_hidden_layers, nodes_per_layer,
loss_fn,num_tot_TX,num_beams)
# Model running/testing
DL_Result={}
mse_list = []
for id in range(0,num_tot_TX,1):
beams_predicted=AP_models[id].predict( In_test, batch_size=10, verbose=0)
DL_Result['TX'+str(id+1)+'Pred_Beams']=beams_predicted
DL_Result['TX'+str(id+1)+'Opt_Beams']=Out_test[:,id*num_beams:(id+1)*num_beams]
mse = mean_squared_error(Out_test[:,id*num_beams:(id+1)*num_beams],beams_predicted)
mse_list.append(mse)
print('mse:',np.mean(mse_list))
DL_Result['user_index']=test_index
mse: 0.00030647155448710575
Attack the RF Beamforming Codeword Prediction Model
The RF beamforming prediction model’s performance is quite good which is 0.00039. Now, we will attack the model with FGSM.
import warnings
warnings.filterwarnings('ignore')
# Model running/testing
DL_Result={}
mse_list = []
for id in range(0,num_tot_TX,1):
# !!!!! Attack generation !!!!
In_test_adv = fgsm(AP_models[id], In_test,Out_test,num_tot_TX,num_beams,eps)
beams_predicted=AP_models[id].predict( In_test_adv, batch_size=10, verbose=0)
DL_Result['TX'+str(id+1)+'Pred_Beams']=beams_predicted
DL_Result['TX'+str(id+1)+'Opt_Beams']=Out_test[:,id*num_beams:(id+1)*num_beams]
mse = mean_squared_error(Out_test[:,id*num_beams:(id+1)*num_beams],beams_predicted)
mse_list.append(mse)
print('mse:',np.mean(mse_list))
mse: 0.0151878763143215
Adversarial Training
The attacked model’s performance (i.e. mse=0.01532) is worst than the normal performance (i.e. mse=0.00039). Let’s train the normal model with adversarial training.
# Model training function
def adv_train(In_train, Out_train, In_test, Out_test,
nb_epoch, batch_size,dr,
num_hidden_layers, nodes_per_layer,
loss_fn,n_BS,n_beams, eps):
in_shp = list(In_train.shape[1:])
AP_models = []
mcp_save = ModelCheckpoint('model.hdf5', save_best_only=True, verbose=0,
monitor='val_mean_squared_error', mode='min')
for idx in range(0, n_BS*n_beams-2, n_beams):
idx_str = str(idx / n_beams + 1)
act_func = 'relu'
model = Sequential()
model.add(Dense(100, input_dim=in_shp[0], activation=act_func))
model.add(Dense(100, activation=act_func))
model.add(Dense(100, activation=act_func))
model.add(Dense(100, activation=act_func))
model.add(Dense(n_beams, activation=act_func))
model.compile(loss=loss_fn, optimizer='rmsprop', metrics=['mean_squared_error'])
history = model.fit(In_train,
Out_train[:, idx:idx + n_beams],
batch_size=batch_size,
epochs=nb_epoch,
verbose=0,
validation_data=(In_test, Out_test[:,idx:idx + n_beams]))
callbacks = [mcp_save]
for _ in range(10):
In_train_adv = fgsm(model, In_train,Out_train, n_BS, n_beams)
In_train_adv = np.concatenate((In_train, In_train_adv), axis=0)
Out_train_adv = np.concatenate((Out_train, Out_train), axis=0)
history = model.fit(In_train_adv,
Out_train_adv[:, idx:idx + n_beams],
batch_size=batch_size,
epochs=nb_epoch*3,
verbose=0,
callbacks=callbacks,
validation_data=(In_test, Out_test[:,idx:idx + n_beams]))
model.load_weights('model.hdf5')
AP_models.append(model)
return AP_models
AP_models = adv_train(In_train, Out_train, In_test, Out_test,
nb_epoch, batch_size,dr,
num_hidden_layers, nodes_per_layer,
loss_fn,num_tot_TX,num_beams,eps)
# Model running/testing
DL_Result={}
mse_list = []
for id in range(0,num_tot_TX,1):
# !!!!! Attack generation !!!!
In_test_adv = fgsm(AP_models[id], In_test,Out_test,num_tot_TX,num_beams,eps)
beams_predicted=AP_models[id].predict( In_test_adv, batch_size=10, verbose=0)
DL_Result['TX'+str(id+1)+'Pred_Beams']=beams_predicted
DL_Result['TX'+str(id+1)+'Opt_Beams']=Out_test[:,id*num_beams:(id+1)*num_beams]
mse = mean_squared_error(Out_test[:,id*num_beams:(id+1)*num_beams],beams_predicted)
mse_list.append(mse)
print('mse:',np.mean(mse_list))
mse: 0.0006801871943259285
All results
Now the RF beamforming codeword prediction model’s performance under attack is better than the undefended model performance.
vals = [0.00039551136731949614,0.01532598592375935,0.0007175213740754419]
df = pd.DataFrame({'Scenario':['SC1: Normal','SC2: Attacked','SC3: Defended'],'vals':vals})
df.plot.bar(x='Scenario',y='vals', rot=0,color=['green', 'red','blue'],
edgecolor='black',legend='',linewidth=2)
plt.xlabel('')
ax = plt.gca()
for i, v in enumerate(vals):
ax.text(-0.17+i*0.97, 0.0002 + v, str(np.round(vals[i],5)), color='blue')
plt.ylabel('MSE')
plt.grid(True, which="both", ls="--")
plt.show()

6. Conclusion
This research discussed one of the security issues related to RF beamforming codeword prediction models’ vulnerabilities and solutions: (1) Is the deep learning-based RF beamforming codeword predictor vulnerable for adversarial machine learning attacks? (2) Is the iterative adversarial training approach a mitigation method for the adversarial attacks in beamforming prediction? We conducted experiments with the DeepMIMO ray tracing scenario to answer these questions. Our results confirm that the original model is vulnerable to a modified FGSM attack. One of the mitigation methods is the iterative adversarial training approach. Our empirical results also show that iterative adversarial training successfully increases the RF beamforming prediction performance and creates a more accurate predictor, suggesting that the strategy can improve the predictor’s performance.