WiFi-Enabled User Authentication through Deep Learning in Daily Activities WiFi-Enabled User Authentication through Deep Learning in Daily Activities

Uniqueness of Individual Characteristics. The distortions of WiFi CSI could reflect a person's minute body movements. Additionally, as demonstrated in Section 3, the amplitude and relative phase in WiFi CSI could be affected by the users’ physiological (e.g., shape and height) and behavioral characteristics (e.g., walking gait, gesture preferences). The system needs to extract effective features from WiFi CSI of daily activities to quantize such unique characteristics of each individual user.

System Robustness and Generality. The collected CSI measurements from real-world environments are usually noisy due to the continuous environmental changes, radio interference, and the like. Besides, the movements of surrounding people could also distort the WiFi CSI, introducing variations in the derived users’ characteristics. Therefore, the system should be robust to capture distinguishable characteristics among users from noisy channel measurements while mitigating the impacts of surrounding people.

Recognizing Activity and Identity Simultaneously. Recognizing activity and user identity simultaneously is very important in many smart home/office enabled applications. For instance, the system can prohibit a specific user (e.g., child) to watch TV at a specific time period. However, recognizing activities and identifying users require a different granularity of features extracted from their activities.

We study the following possible attacks that might be harmful to the proposed authentication system.

• Naive Attack. The attacker does not have prior knowledge about the activities the legitimate users performed. In order to pass authentication, the adversary attempts to conduct random activities to create similar impacts on the WiFi signals as the legitimate user.
  
• Content-Aware Attack. The attacker knows about which types of the activities are used for user authentication, but has not observed how the legitimate user performs them. The adversary tries to pass the authentication through performing the same activities as the legitimate user.
  
• Knowledgeable Observer Attack. The attacker is capable of observing the activities performed by the legitimate user for whom the proposed system is authenticating via shoulder surfing or videotaping. The adversary tries to perform the same activity and imitate the legitimate user's behaviors to pass the authentication.
  

The basic idea of our system is to capture the unique physiological and behavioral characteristics inherited from human daily activities for user authentication leveraging WiFi signals. The users have habitual patterns on their behaviors, so the daily activities usually present high consistency for each individual [26]. As illustrated in Figure 4, our system takes as input CSI measurements from WiFi links between WiFi-enabled IoT devices (e.g., smart appliances), and then extracts both the CSI amplitude and the relative phase for each OFDM subcarrier for signal pre-processing. Unlike previous studies [30, 35, 36] which only utilize CSI amplitude, we explore relative phase along with CSI amplitude to capture representative characteristics through both large-scale walking activities and small-scale stationary activities. Given the amplitude and relative phase information, a band-pass filter is first deployed to eliminate the environmental interferences (e.g., reflected signals from furniture and walls) and ambient noises. We also propose a subcarrier selection algorithm to pick out the subcarriers with stable CSI measurements, which could represent reliable activity characteristics. Before performing features extraction, we examine the moving variance and related short time energy (STE) of the pre-processed data to determine the CSI segments, which capture the location changes for walking activities and body movement for stationary activities.

Next, we will present the core components of our system, Physiological and Behavioral Feature Extraction and Deep Learning-Based User Authentication. We perform activity recognition and user authentication based on the physiological and behavioral features extracted from CSI measurements, which characterize both human activity/identity uniqueness. The system extracts six time domain and three frequency domain features to capture both the physiological and behavioral characteristics of users such as height, shape, and behavioral preference. Specifically, the time domain features, including maximum, minimum, mean, skewness, kurtosis, and standard deviation, aiming to represent the extent of human movements and contour of human body, while the frequency domain features, including spectrogram energy, percentile frequency component, and spectrogram energy difference, are used to depict the fine-grained behavioral characteristics such as moving speed of torso and leg. All the above CSI-based features together provide a comprehensive and detailed representation for walking/stationary activities.

Finally, our system performs activity recognition and human authentication by building a three-layer DNN model based on both AutoEncoder [6] and CNN architectures. To mitigate the interferences from surrounding people, we propose to explore the spatial diversity benefit from MIMO technology and design a CNN-based model to achieve more reliable authentication. Unlike previous authentication schemes based on high-dimension feature sets and linear classification models (e.g., SVM), our DNN model learns non-linear biometric abstractions which are computation efficient and are robust to small-scale input variations (e.g., the variations of features caused by the wearing changes of users). Particularly, we obtain the biometric abstractions with respect to single activity and authenticate the user based on the corresponding CSI activity segment. Figure 4 illustrates the functionality of each layer in our deep learning architecture for people authentication. In particular, the first level coarsely distinguishes the activity types (i.e., walking or stationary activity); the second layer exploits deep representations of the first layer and obtains the activity details such as walking trajectories and detailed stationary activity types (e.g., turning on a light); and the third level obtains even deeper representation of the features and finally completes user authentication process. To reduce the profiling efforts, we design a transfer-learning-based mechanism to adapt the DNN model to new users and environments by retraining the model with a reduced profile size. Additionally, our system is resilient to user spoofing, who either does not exist in legitimate user profiles or tries to mimic a legitimate user's activity, by using a SVM-based model with the generated DNN abstractions.

To ensure the reliability of the features extracted from the CSI measurements, data calibration and subcarrier selection techniques are developed to mitigate the ambient noises and select the subcarriers with stable CSI measurements, respectively. The details are presented in Section 7.

Both walking and stationary activities lead to the variations in wireless channel, resulting in changing CSI measurements. So we apply the short time energy (STE) upon CSI amplitude's moving variance to detect human activities, and then perform corresponding data segmentation. Moreover, stationary activities (e.g., opening a cabinet) usually involve relative smaller-scale body movements than walking activities, which makes them even harder to be detected. We thus propose to examine STE which is more sensitive to subtle body movements because it is a summation of squared signals within a sliding window. We calculate STE within a sliding window as follows:

Figure 5(a) shows both CSI's moving variance and STE for two rounds of the same stationary activity (i.e., opening a cabinet). When the activity occurs, we can observe that STE exhibits a greater magnitude, showing its potential at activity detection. Furthermore, we also found the peaks of the fluctuating part in STE always locate around the center of the activity duration. We are thus inspired to utilize a dynamic threshold, which can be applied to all types of activities, to perform activity detection and corresponding data segmentation. Specifically, a weight $w=0.1$ derived from our empirical study is deployed for the dynamic threshold calculation: $\tau =w*E$, where $E$ is the maximum value of STE for an individual activity. We then search for the starting and ending points, $t_s$ and $t_e$, of this activity by solving the following objective problem:

where $t_m$ is an arbitrary time index in the middle of activity duration. Figures 5(b) and 5(c) show the segmented time series of CSI amplitude and relative phase of the $\text{first}$ subcarrier during two rounds of the stationary activity. The results demonstrate the efficiency of our activity detection/segmentation algorithm.

To capture the unique physiological and behavioral characteristics inherited from users’ daily activities, it is essential to extract effective and reliable features from the CSI measurements. In particular, both time and frequency domain features based on CSI amplitude and relative phase information are examined to discriminate different users.

Time Domain Feature Extraction. In our system, six time domain features with respect to CSI amplitude and relative phase, including maximum, minimum, mean, skewness, kurtosis and standard deviation, will be extracted to characterize both human activity and identity uniqueness. In order to preserve temporal patterns of the user activity and provide finer feature granularity, we first partition the CSI segment into $l$ chunks of equal length, and then extract the six time domain features from each chunk. We empirically set $l$ as 20, which provides 120 feature points from each subcarrier. Figure 6 and 7 present the extracted time domain features for the same stationary activity (i.e., opening a cabinet) performed by two users based on amplitude and relative phase, respectively. We can find that these features are significantly different between two users. It encourages us to leverage these time domain features to capture human unique characteristics inherited from their daily activities.

Frequency Domain Feature Extraction. As indicated in previous work [31], CSI measurements in the frequency domain are able to reveal the speeds of WiFi path length changes caused by human movements. Therefore, besides the time domain features, we also extract the representative features in frequency domain to capture the users’ behavioral characteristics (e.g., walking gait, gesture preference). Given a CSI segment, we first adopt the short-time Fourier transform (STFT) to obtain the two-dimensional spectrogram for the CSI amplitude and relative phase of each subcarrier. More specifically, we calculate 1,000 points FFT within a $100\text{-ms}$ sliding window, shifting 50 ms each time. We then use bicubic interpolation [10] to resize the spectrogram into a matrix $M_{(i,j)}$ (i.e., 10-by-10 matrix) of fixed size, which maintains spectrogram in a consistent feature space for different activities. Next, three frequency domain features on the top of $M_{(i,j)}$ are extracted: (1) Spectrogram Magnitude. Each element in the matrix $M_{(i,j)}$; (2) Percentile Frequency Components (PFC). $PFC(i,n)=\frac{\sum _{j=1}^{n}M_{(i,j)}}{\sum _{j=1}^{10}M_{(i,j)}}$, where $n=1\dots 10$, subjected to and , indicating the moving speed of torso and leg [30]. (3) Spectrogram Difference between Time Windows. The element-wise differences between two consecutive rows in $M_{(i,j)}$, which capture the acceleration or deceleration process of body movement. In total, we extract 210 frequency domain features from the CSI segment with respect to one specific activity.

In many shared spaces (e.g., corporate offices, apartment living rooms), the CSI measurements to capture a user's physiological/behavioral characteristics are easily distorted by the movements of surrounding people. We assume that the surrounding people keep a proper distance from the target user, so that propagation paths could capture the user's activity while remaining unaffected by the surrounding people. To effectively capture such propagation paths, we explore the spatial diversity benefit from MIMO technology and propose a CNN-based model to achieve more reliable user authentication.

Given the high-dimensional features extracted from multiple wireless antennas, the autoencoder architecture proposed in our previous work is not effective. This is because its fully connected network structure requires a huge number of neurons for processing a large set of features. Additionally, such architecture does not take into account the spatial diversity of the feature set from multiple antennas, treating the input features of different antenna pairs as those of single antenna pair. Therefore, we exploit CNN [13], a more effective deep-learning framework, for multi-antenna abstraction extraction.

The proposed CNN model could constructively integrate features from multiple antennas and extract feature abstractions that are more robust under the impacts of surrounding people. It first packs the feature vectors from all $M$ available antenna pairs into a matrix $C=\lbrace C_1;...;C_M\rbrace$. Then, the CNN model processes the data matrix leveraging three stacked hidden layers and extract different levels of abstractions for activity type recognition, activity recognition, and user authentication, respectively. Each neuron applies convolution operation (i.e., cross-correlation) to a subset of features/abstractions, instead of the whole feature/abstraction set. In this way, the CNN model is able to process a large feature set with fewer neurons while extracting effective multi-antenna feature abstractions.

Each hidden layer of the CNN consists of three components, a 2D convolutional layer, a rectified linear unit (ReLU), and a pooling layer. The 2D convolution layer contains a group of $K$ neurons, where each neuron acts as a filter that iteratively computes dot products between a learnable coefficient matrix $W^k$ with height $h$ and width $w$, and subsets of inputs. Specifically, the input subsets are obtained by applying a 2D sliding window on the layer input. The sliding window has a step size of 2 and same height and width as $W^k$. By setting $h$ as the input height (e.g., number of antennas, $M$), the sliding window could include a subset of features/abstractions covering all available antenna pairs. Thus, the derived abstractions are capable to represent all the physiological and behavioral characteristics of users captured by different antenna pairs. We empirically set the width of $w$ as 3. Then a ReLu layer is attached to the convolutional layer to introduce nonlinearity by replacing negative neuron outputs with 0. A max-pooling layer is added to reduce the size of abstraction vector in each hidden layer. The hidden layers are trained with stochastic gradient descent with momentum (SGDM) optimizer [24] with a learning rate of 0.01. The structure of CNN could facilitate abstraction extraction given high-dimensional inputs.

Given three hidden layers with either autoencoder or CNN architecture, we stack the hidden layers and a softmax layer to construct a hierarchical model for activity recognition and user authentication. Previous work [28] found that higher level feature abstractions are more robust to small-scale input variations, which meets the hierarchical requirements of our system. Additionally, the three-layer DNN itself can only derive compressed representations of physiological and behavioral characteristics, so we still need a softmax function [2] in each layer to complete the activity recognition and user authentication process in a hierarchical order. Specifically, we define the softmax function as follows:

Our system constructs a CNN model consisting of three softmax layers for activity separation, activity recognition, and user authentication, respectively. During the user enrollment phase, we collect activity data from each legitimate user as the training data (i.e., user profile) for model construction. The training data are segmented and processed as discussed in Section 5. To reduce the profiling efforts, we find a minimal profile size (i.e., the number of activity segments) for each individual user each activity. Our strategy is to examine the user identification accuracy during data collection. In particular, our system retrains the CNN model after collecting every activity segment from the user and classifies the identity. The data collection process would stop when the user identification accuracy meets a predefined threshold (e.g., $90\%$). The system will profile other activities of the user in the same manner. By using this strategy, our system can find a minimal profile size for each user. We will discuss the activity recognition/user authentication accuracies against the data size for determining the threshold in Section 8.5.

After model construction, the CNN model may still need to update due to many practical factors, such as new enrollments and environment changes. To reduce the training efforts when updating the CNN model, we introduce a supervised transfer learning-based approach, which could adapt the previously trained CNN model based on existing data for new users or new environments. Please note that we only apply transfer learning to the proposed CNN architecture. We exploit inductive transfer learning [19], where the settings of the CNN model (i.e., the number of layers and neurons) are the same as the original one, while the number of categories (i.e., users/activities) is increased with new enrollments. The parameters of the new CNN model is initialized with the CNN trained with existing data. Next, we attach a softmax layer with extended categories to the new CNN model for classifying a larger group of users, including new users. To accommodate new users/activities in our system, the model and the softmax layer parameters, $\theta _n$, will be fine-tuned as following:

Besides adapting the deep learning model to new users and environments, our transfer learning technique can also accommodate the users’ biometric variations associated with age changes, which result in different weight, height, and behaviors. Our system can retrain the CNN model with only a few new activity segments of an enrolled user, which adapts the model to identify the user with changed biometrics. For adolescents with rapid growth in height and varying behaviors, we will keep updating their profiles by storing activity segments during authentication and retrain the model. Compared to new user enrollment and environment changes, updating the biometric profile of an enrolled user would be less challenging since the user's physiological and behavioral characteristics are partially shared in the trained model.

Besides three-layer DNN, we also adopt a SVM model [3] to determine whether the activity the user performed matches one of the legitimate user profiles. Particularly, we propose to utilize one-class support vector machine with Gaussian kernel for detecting the user spoofing, who either does not exist in legitimate user profiles or tries to mimic a legitimate user's activity. We first construct an one-class SVM model for each of the legitimate users based on the high-level abstractions from the DNN model. We then derive a class score $S_u$, which compares the similarity between the feature abstractions of each testing sample and that of the user $u$:

To ensure reliable feature extraction, we preprocess the raw CSI measurements with phase unwrapping and a band-pass filtering techniques, which are effective on eliminating the environmental interferences and ambient noises. Specifically, we first eliminate the relative phase error caused by the phase offset on each subcarrier. As shown in Figure 9(a), the raw relative phase at three subcarriers (i.e., subcarriers 4, 5 and 6) have obvious discontinuities between consecutive packets when the relative phase value is close to $\pm \pi$. To eliminate such discontinuities, a $\pm {2\pi }$ is added to the relative phase of the later packet if the absolute phase difference of two consecutive packets is greater than or equal to $\pi$. Figure 9(b) shows the corresponding relative phase streams after phase calibration.

Besides the phase offset, the amplitude and relative phase in CSI measurements are also easily affected by the low-frequency interference (i.e., caused by signals reflected from static objects) and high-frequency noise. In order to eliminate the above impacts while preserving the user physiological and behavioral characteristics in CSI measurements, a bandpass Butterworth filter [20] is adopted. Previous work [31] found that the frequency range of most human activities including running in a fast speed exhibit CSI frequency components less than $300 \text{Hz}$. We thus adopt a relative low frequency band-pass Butterworth filter (i.e., with passing band $5 \text{Hz}$–$100 \text{Hz}$) to effectively remove both low- and high-frequency components from the spectrum. Given the example scenario where a person walks in a room between 2 and 7 seconds, Figure 10(a) shows the spectrogram of the corresponding time series of CSI amplitude at a subcarrier (i.e., subcarrier 1). We can observe that the spectrogram exhibits extremely high energy level in the low-frequency band (i.e., $\lt 10 \text{Hz}$) even the person remains static. As the spectrogram after band pass filtering shown in Figure 10(b), we can observe that the CSI amplitude pattern caused by human walking is still preserved while irrelevant frequency components are removed.

Our preliminary study finds that the CSI measurements of several subcarriers are more sensitive to ambient noise. To ensure the reliability of CSI measurements for later processing, we propose a new subcarrier selection method to determine the noise-resilient subcarriers from the CSI measurements. The CSI measurements at neighboring subcarriers are usually highly correlated, however such correlation could be destroyed by heavy noises on some of the subcarriers. To eliminate the negative effects caused by the unstable subcarriers, a covariance-based scoring function is defined to assess each subcarrier's correlation level with its neighboring subcarriers as follows:

Devices and Network. We emulate the WiFi network in IoT environments with two commercial laptops equipped with $802.11n$ WiFi NICs (i.e., Intel 5300 NICs). Specifically, we deploy a Dell $\text{E}6430$ laptop as transmitter and a Lenovo $\text{T}61$ laptop as receiver. Both the transmitter and receiver are equipped with three embedded antennas and run Ubuntu 14.04 operating system with the 4.2.0 kernel for measuring CSI over 30 subcarrier groups [8]. We extract the CSI amplitude on the link between the main antenna pair (i.e., 1st antenna in both transmitter and receiver), and compute the relative phase of CSI between the two links from the transmitter's main antenna to the first two antennas on the receiver. For multi-antenna abstraction extraction, we extract CSI amplitude/relative phase from all transmitter-receiver antenna pairs (i.e., 9 links). The packet transmission rate is fixed at $1,000$ pkts/s to enable high-resolution analysis in the frequency domain.

Evironments and Activities. The proposed system is evaluated in both a university office and an apartment with the size of $26 \text{ft}\times 14 \text{ft}$ and $36 \text{ft}\times 22 \text{ft}$, respectively. Figure 12 shows the experimental setups involving two laptops to emulate as IoT-enabled devices (e.g., smart refrigerator and smart TV) and generate WiFi traffics. A total of 8 walking activities and 8 stationary activities (30 rounds for each) are performed by 11 and 5 volunteers in these two indoor environments, respectively. Fourteen volunteers are males and the other two volunteers are females. The weights of the 16 volunteers are within 53 kg—78 kg, and their heights range from 1.51 m—1.87 m. Some volunteers have similar body shape and height. Due to the functionality differences of the two environments, we choose different yet still typical stationary activities in the two environments. The details of the activities are listed in Table 1, and the locations of stationary activities and walking trajectories are also shown in Figure 12. In total, we collect $3,336$ activity segments performed by 11 subjects in the office environment, and 834 activity segments performed by 5 subjects in the apartment. Unless mentioned otherwise, half of the collected dataset (i.e., 15 rounds of each activity per subject) is used for training the DNN model, and the rest of data is used for testing the system performance.

Classification Strategies. In the user enrollment phase (i.e., training phase), our system collects CSI measurements through WiFi scanning while people are performing daily activities. Then, we associate activity/identity labels with the corresponding CSI segments as the training data (i.e., user profile). We extract both time and frequency domain features from the data and feed them to the CNN model for training. We empirically determine the filter size of each CNN layer as 64 and kernel size as $(3, 3)$. During the testing phase, our system first segments and processes the CSI measurements, which is the same as that in the training phase, and then identifies the user through the trained CNN model. The inference time for an activity segment is around 0.78 seconds (i.e., including feature extraction and user identification).

Evaluation Metrics. To evaluate our system, we use four evaluation metrics: identification accuracy, confusion matrix, true-positive rate (TPR), and false-positive rate (FPR). Following existing work on using WiFi signals for user identification [35], we use identification accuracy as the main evaluation metric. It is defined as the percentage of the human identity/activity correctly recognized by our system. We use the confusion matrix to further illustrate user identification/activity recognition performance, with each entry represents the percentage of correctly classified identify/activity. Each column in the confusion matrix indicates the ground truth of an identity/activity and each row represents the classified identity/activity in our system. TPR denotes the percentage of users that are correctly verified as legitimate users, and FPR is the percentage of attackers that mistakenly pass our system.

We first present the performance of the proposed system on user identification in both office and apartment environments. As shown in Figure 13(a), we observe that, our system achieves over $93.0\%$ user identification accuracy for all users and the average accuracy is $96.8\%$ with a standard deviation of $2.6\%$ in the office environment. Compared with WiWho [35], our system can identify a larger group of people (i.e., 11 vs. 6) with higher user identification accuracy (i.e., $96.8\%$ vs. $80.0\%$). Figure 13(b) gives the confusion matrix for the user identification in the apartment. Our system achieves over $93.0\%$ identification accuracy for four of the users. The average user identification accuracy is $95.4\%$ with a standard deviation of $1.8\%$. We have comparable high accuracies on user authentication in the two different environments, and thereby confirm the effectiveness and reliability of the proposed system on user identification.

Robustness to Spoofing Attacks. As indicated in Section 6.4, threshold selection is critical for accurate detection of spoofing attacks. To obtain an appropriate threshold, we simulate a spoofing attack scenario in an office environment, where 8 of the 11 users act as spoofers and the rest are legitimate users. A threshold that maximizes the TPR while minimizing the FPR is selected for evaluating our system under naive, content-aware, and knowledge observer attacks.

We take turns to set each participant as the legitimate user and the remaining users as the attackers in both office and apartment environments. We first evaluate the performance of our system on defending against naive attacks, where the attacker does not possess prior knowledge about the user's activities. As shown in Figure 14, our system can achieve over $96.3\%$ TPR with less than $3.2\%$ FPR for all users in the two environments. The results confirm that the random activities of attackers can hardly create similar biometrics as those of the profiled activities of the legitimate users, and thus the proposed system can reliably defend against random attacks. Figure 15 shows the performance of the proposed system under content-aware attacks, where the attackers try to pass the authentication system through performing the same activity as the legitimate user. The average TPRs are $88.5\%$ and $90.4\%$ for the office and the apartment environments, respectively, with FPR lower than $10\%$ in both environments. The results show that even when the adversary has a similar body shape and height to the legitimate user, our system can still identify the legitimate user based on unique behavioral characteristics. Figure 16 shows the robustness of our system against the knowledge observer attack, the most extreme attack, where the attacker is capable of observing the behaviors of the legitimate user. For both environments, we find our system can still achieve over $86.3\%$ and $87.6\%$ TPR, respectively. The FPRs are less than $8.3\%$ and $10.1\%$ in the two environments. The results show that the proposed system is still effective in defending against the knowledgable observer attacks.

We further analyze the system's robustness using entropy-based metric, which quantifies the system's security strength. Due to the limited size of our dataset, we use the Euclidean distance between every pair of feature abstractions from different users to calculate the entropy, instead of performing uncertainty analysis on a large group of users. The entropies for the spoofing detection in the two environments are 7.9 and 7.4, respectively. This means that our system at least has a similar security level to human chosen 4-digit PINs [9].

We examine the activity recognition performance in the DNN model. Figure 17 depicts the confusion matrix for activity recognition (i.e., outputs of DNN layer 2) in both office and apartment environments. The average activity recognition accuracies for both stationary and walking activities are as high as $98.6\%$ and $99.1\%$ in the office and apartment, respectively. Further, it is encouraging to find that DNN model achieves similar accuracy for both stationary and walking activities. The slight difference on the recognition accuracy between the two types of activities is caused by the limited resolution of WiFi signals on capturing small scale body movements for stationary activities. Overall, the proposed DNN model is highly effective on recognizing different types of activities.

Impact of Surrounding People. In the office environment, we ask one, two, or three participants to act as surrounding people and to walk around when a target subject is conducting activities, which is a typical scenario in small offices and homes. We use walking activities to examine the system's robustness since they produce more significant interference to CSI than stationary activities (e.g., turning on a light, sitting). In the experiment, the participants walk 6 feet away from the target subject, without cutting the LOS path between the transmitter and the receiver, and the distance between every two people is around 6 feet . We do not limit their walking trajectory/speed. We recruit another five subjects to act as target subjects and each of them conducts six walking and four stationary activities. The activity profiles are constructed by using the CSI measurements collected without the interferences. Figure 18 shows the performance of multi-antenna- and single-antenna-based abstraction extraction schemes. We can observe that the multi-antenna based-scheme could help to maintain high user recognition/authentication accuracies under interferences from different number of surrounding people. As shown in Figure 18(a), the proposed scheme achieves $93.1\%$, $90.2\%,$ and $87.1\%$ average accuracies under the interference of 1, 2, or 3 surrounding people, respectively. We can also find that the single-antenna-based scheme is susceptible to the impacts from the activity of surrounding people. Furthermore, as shown in Figure 18(b), the multi-antenna-based scheme could also maintain a high activity recognition accuracy under the impacts of people moving nearby. The results confirm the effectiveness on mitigating the interferences of surrounding people.

Feature Comparison. To further analyze the impact of different features on the system performance, we compare the authentication performance using different kinds of CSI features in both time and frequency domains: Amplitude, Relative Phase, and all of these features (i.e., Combined). We present the comparison results of user authentication accuracy in Figure 19 for both office and apartment environments. Figure 19(a) shows that the CSI relative phase features have relative higher user identification accuracies for stationary activities (e.g., $a,b,c,d$) comparing to the amplitude feature. This is primarily because relative phase exhibits higher sensitivity on capturing small-scale body movements. In addition, we also find in both Figures 19(a) and 19(b) that the combined features of both CSI amplitude and relative phase achieve the best performance, indicating the combining features can provide the finest features to distinguish individual subjects.

Impact of Training Size. The DNN model needs to build CSI profiles for each activity or each individual before performing activity recognition and human authentication. It is necessary to study the impact of training size on system performance. Here we define the training size as the number of training samples for each activity or for each individual. As shown in both Figures 20(a) and 20(b), our system can achieve consistently high accuracy on user identification and activity recognition with different training sizes. Especially, our deep learning model maintains over $90\%$ accuracy on user identification and activity recognition even with the training size as small as 4. The above results show that our system has minimum requirements on building the CSI profile while ensuring remarkable performance.

Impact of Sampling Rate. In order to validate that our user authentication scheme can work under various sampling frequencies of WiFi-enabled IoT devices, we evaluate our system under different frame rates. We show the average user authentication accuracy of office and apartment under different sampling rates in Figure 21. We can observe that our system can maintain high accuracy across different frame rates from $200 \text{Hz}$ to $1,000 \text{Hz}$. Particularly, the authentication accuracy is still over $86\%$ even for the low sampling rates such as $200 \text{Hz}$ and $400 \text{Hz}$. The above observations confirm that our system can be applied on IoT devices with different sampling capabilities.

Finally, we conduct experiments with five new users on evaluating the performance of the proposed transfer learning technique. Specifically, we reuse the parameters of a pre-trained CNN for differentiating five existing users to initialize the weights of a new CNN model for recognizing new users. Additionally, we leverage another CNN model directly trained with data of the same 10 users as the baseline. Both the existing and new CNN models are trained/fine-tuned with all data from the same amount of samples (i.e., 2-14) from new users. As shown in Figure 22(a), transfer learning can greatly improve user identification accuracy under various training sizes for the new users. Especially, our transfer learning method could help to achieve $86.1\%$ accuracy even with the training size as low as 2. Such improvements could greatly reduce training samples required for new enrollments. Our other experiments with fewer new users (e.g., $1\sim {2}$ users) show more significant improvements when utilizing the transfer learning technique.

We also exploit transfer learning to adapt an existing CNN model to new environments. Both the existing and the new CNN models are trained with data from both environments (i.e., apartment and office) but the existing CNN model does not reuse the parameters of the pre-trained CNN. As shown in Figure 22, transfer learning could help to increase the user identification accuracy by around $10\%$ for each of the training sizes. The results show that the proposed transfer learning based scheme could greatly improve system extensibility for new user enrollments/environments.