Digital Audio Tampering Detection Based on Deep Temporal–Spatial Features of Electrical Network Frequency

Abstract

In recent years, digital audio tampering detection methods by extracting audio electrical network frequency (ENF) features have been widely applied. However, most digital audio tampering detection methods based on ENF have the problems of focusing on spatial features only, without effective representation of temporal features, and do not fully exploit the effective information in the shallow ENF features, which leads to low accuracy of audio tamper detection. Therefore, this paper proposes a new method for digital audio tampering detection based on the deep temporal–spatial feature of ENF. To extract the temporal and spatial features of the ENF, firstly, a highly accurate ENF phase sequence is extracted using the first-order Discrete Fourier Transform (DFT), and secondly, different frame processing methods are used to extract the ENF shallow temporal and spatial features for the temporal and spatial information contained in the ENF phase. To fully exploit the effective information in the shallow ENF features, we construct a parallel RDTCN-CNN network model to extract the deep temporal and spatial information by using the processing ability of Residual Dense Temporal Convolutional Network (RDTCN) and Convolutional Neural Network (CNN) for temporal and spatial information, and use the branch attention mechanism to adaptively assign weights to the deep temporal and spatial features to obtain the temporal–spatial feature with greater representational capacity, and finally, adjudicate whether the audio is tampered with by the MLP network. The experimental results show that the method in this paper outperforms the four baseline methods in terms of accuracy and F1-score.

1. Introduction

The boom in Internet information technology has made digital audio, such as telephone recordings, voice messages, and music files, readily available in our daily lives [1,2]. Due to the low threshold and powerful operation of existing audio editing software, digital audio tampering can be easily accomplished by an average user without any expertise in audio processing [3,4]. In addition, millisecond digital audio tampering fragments are often difficult to identify [1,5,6], and unscrupulous individuals may use digital audio tampering to try to evade legal sanctions and even cause harm to society. As a result, digital audio forensic methods are increasingly in demand in areas such as judicial forensics, scientific discovery, and commercial applications to reduce the impact caused by such incidents [7,8,9,10,11,12].

Forensic techniques for digital audio are mainly divided into two types: active forensics and passive forensics [1]. The active forensic technique of digital audio is mainly used to determine the authenticity or integrity of audio by detecting whether the pre-embedded digital signature or digital watermark is corrupted. However, in practical applications, most of the audio signals are not pre-embedded with watermarks or signatures at the time of recording, so digital audio active forensic techniques have limitations in applications. Passive detection of digital audio tampering means that there is no need to add any information, and the authenticity and integrity of digital audio are discriminated only by the characteristics of digital audio itself. Passive detection of digital audio tampering is more practical for audio forensics in complex environments, which is why the method proposed in this paper focuses on it.

In recent years, research results in the field of passive detection of digital audio tampering have focused on the selection of audio features, such as difference information of background noise [13,14,15,16,17], spectrograms of audio content [18,19], pitch [20,21,22], resonance peaks [22], ENF difference information [23], ENF harmonic signals [24], ENF phase and frequency information [6,25,26], etc. The ENF is automatically embedded in the audio when recording and is characterized by random fluctuations at nominal frequencies (50 Hz or 60 Hz) with some stability and uniqueness [27]. Therefore, ENF is widely used for tampering detection of digital audio.

Most of the methods based on the ENF digital audio tampering detection extract ENF feature information and achieve tampering detection by classification algorithms. However, in the selection of features, most researchers only use spatial or temporal information of the ENF, resulting in a certain degree of loss of tampering information, which leads to a weak characterization of features and a low classification accuracy.

To solve the problems of weak feature representation and low classification accuracy, inspired by the success of deep representation learning in speaker recognition [28,29], computer vision [30,31,32,33,34,35,36,37,38], and big data [39,40], this paper proposes a digital audio tampering detection method based on the ENF deep temporal–spatial feature. Moreover, this paper first extracts the phase sequence of the ENF by using the first-order DFT analysis method and then divides the frames according to the ENF temporal variation information to obtain the time series matrix of ENF phases to represent the shallow temporal features of ENF. Finally, the unequal phase sequences are framed by adaptive frameshifting to obtain a matrix of the same size to represent the shallow spatial features of ENF. The construction of the parallel RDTCN-CNN network model mainly consists of four parts: a deep temporal feature extraction module, a deep spatial feature extraction module, a temporal–spatial feature fusion module, and a classification module. In the deep temporal feature extraction module, we extract deep temporal features based on the causal convolution principle of RDTCN. In the deep spatial feature extraction module, we extract deep spatial features by using the excellent spatial representation ability of CNN. In the temporal–spatial feature fusion module, we use the branch attention mechanism to adaptively assign weights to deep spatial and temporal features to obtain the fused temporal–spatial features. In the classification module, we determine whether the audio is tampered with or not by a multilayer perceptron (MLP).

The main contributions made in this paper are as follows:

• Based on the extraction of high-precision ENF phase sequences, we frame the ENF according to its temporal volatility variation to represent the temporal features of the ENF and frame the ENF by adaptive frameshifting to obtain a phase feature matrix of the same size to represent the spatial features of the ENF. The feature representation capability is enhanced by deeply mining the tampering information of different dimensions in the ENF.
• We exploit the excellent modeling ability of RDTCN in the time domain and the spatial representation ability of CNN to extract deep temporal and spatial features and use the branch attention mechanism to adaptively assign the weights of temporal and spatial information to achieve the fusion of temporal–spatial features. The fused temporal–spatial features, with complementary advantages, are beneficial to improve detection accuracy. The implementation code for this study is posted at https://github.com/CCNUZFW/DTSF-ENF (accessed on 21 February 2023).
• The proposed framework achieves state-of-the-art performance on the datasets Carioca, New Spanish, and ENF_Audio compared to the four baseline methods. Compared with the baseline model, the accuracy is improved by 0.80% to 7.51% and the F1-score is improved by 0.86% to 7.53%.

The rest of this paper is organized as follows. Section 2 describes the existing related work. In Section 3, we provide the problem definition for this study and summarize the important symbols that appear in this paper. Section 4 describes the framework proposed in this paper. Section 5 presents the dataset used to evaluate the performance of the framework, details of the specific experimental setup, and comparison experiments. Finally, the paper concludes in Section 6 and lists the directions for future work.

2. Related Work

Digital audio files are obtained by recording equipment in a certain environment, so the audio must contain features of recording equipment and recording environment, and these features have a certain degree of stability. From the perspective of digital audio components, these features can be divided into three kinds: (1) based on background noise consistency detection; (2) based on the analysis of audio content features; (3) based on electrical network frequency consistency detection.

2.1. Based on Background Noise Consistency Detection

Digital audio is recorded in a specific environment, and when audio is recorded in a complex environment, the recorded audio will contain background noise information in the current environment. When audio tampering operations such as deletion and insertion occur, it will lead to discontinuity of the background noise in the audio, so the detection of tampering operations can be performed by analyzing the background noise part of the digital audio [41].

When recording audio in a room or closed environment, reverberation is introduced in the recorded audio. Many scholars have conducted audio forensic studies through the reverberation of audio. Malik et al. [42] analyzed the consistency of reverberation in digital audio with the recording environment by statistical methods. Mascia et al. [43] classified the recording environment by using the reverberation time and Mel Frequency Cepstral Coefficient (MFCC) as features. The consistency of the background noise when recording audio in a noisy environment can also be used as a basis for determining whether the audio has been tampered with. Noise information is separated from digital audio and its consistency is analyzed for audio tampering detection. Ikram et al. [44] proposed to first extract the preliminary noise signal based on the spectral estimation of geometric transformation, and then remove the speech residual part of the preliminary noise signal using the audio higher-order harmonic structure feature to obtain a purer noise signal. To detect audio splicing operations using information such as background noise in the audio, Meng et al. [13] detected heterogenous splicing tampering of audio by comparing the similarity between the background noise variance of syllables.

Although several research results have been obtained based on the analysis of background noise, there are still some limitations. One is that the complexity of the actual environment is difficult to predict, and the second is that the audio noise separation under very short noise samples is also difficult to carry out. In addition, in the consistency analysis of noise or reverberation, how to select the optimal feature set to characterize the environmental noise is still a problem worth exploring.

2.2. Based on the Analysis of Audio Content Features

The tampering of audio speech content leads to the weakened and abrupt inter -frame correlation of audio characteristics, so we can determine whether digital audio has been tampered with based on the variability of audio content features [14,22]. Chen et al. [45] implemented the detection of tampered audio in the time domain by performing discrete wavelet packet decomposition and singularity analysis on the speech signal. Imran et al. [46] used chaos theory so that tampering points may exist anywhere in the audio, and then detected copy-paste tampering by comparing the differences in the speech spectrogram of the turbid parts. Yan et al. [21] used the sequence of fundamental and resonant peaks of the turbid segment of the audio as features and achieved copy-paste tampering detection by comparing the similarity with a threshold value, and the method is highly robust to common post-processing tampering operations.

After most tampering operations are performed, some post-processing operations are often performed to mask the tampering traces [19]. Therefore, when such post-processing operations are detected in the audio, this audio may have been edited. Yan et al. [3] detected the smoothing operations of editing software using a support vector machine (SVM) based on the local variance of the differential signal.

2.3. Based on Electrical Network Frequency Consistency Detection

When the recording device is powered by the electrical network, the ENF is automatically embedded in the recording file, and because the ENF signal has a certain stability and uniqueness [25] and shows reliable discrimination in audio tampering detection studies, the technique based on ENF analysis has been widely used in the field of digital audio tampering detection, and it is also one of the most effective methods in the last decade [1].

Esquef et al. [47] proposed the TPSW method to estimate the level of ENF background variation based on the fact that the tampering operation causes a sudden change in the ENF instantaneous frequency at the tampering point, using the Hilbert transform to calculate the instantaneous frequency, thus obtaining the mutation point as both the tampering operation point, and the accuracy of the algorithm is better than that of the Rodríguez method [25]. In addition, Reis et al. [26] proposed an ESPRIT-based estimation of the phase peak feature to measure the fluctuation of ENF for the case of phase discontinuity of the tampered signal ENF and used SVM to automatically detect the abrupt change of ENF. However, using only the tampering information in the phase features of ENF has some limitations in the feature characterization capability. On this basis, Wang et al. [6] extracted the phase features of the ENF component (ENFC) based on $DF{T}^0$ and $DF{T}^1$, extracted the instantaneous frequency features of ENFC based on Hilbert transform, and used the SVM classifier to determine whether the signal has been tampered with or not. In addition to using ENF phase and frequency features, Mao et al. [48] extracted ENF features from audio signals using multisignal classification, Hilbert linear prediction, and the Welch algorithm and detected the extracted features by convolutional neural networks. Sarkar et al. [49] decomposed the extracted ENF into low and high outlier frequency segments and then used statistical and signal processing analysis to determine the potential feature vectors of the ENF segments, and finally, an SVM classifier was used for validation.

To further increase the detection accuracy and robustness of ENF-based audio tampering methods, some researchers have investigated the characteristics of ENF to obtain better-quality features. Karantaidis et al. [50] added a customized delay window to the Blackman–Tukey acoustic spectrum estimation method in order to reduce the interference of speech content to make the estimated ENF with higher accuracy. There are often some noises and interferences in the audio, and Hua et al. [51] proposed a Robust Filtering Algorithm (RFA) to enhance the ENF signal in the audio, and this method makes the extracted ENF signal more accurate. Based on this, Hua et al. [24] used RFA to enhance each harmonic component of ENF and finally combined the harmonic components in a weighted manner to finally obtain a more accurate ENF estimation.

For feature selection, tampering detection of audio is usually achieved by extracting ENF static spatial information. However, these methods cause the loss of ENF temporal information to the extent that the feature representation is weak. In classification algorithms, audiovisual analysis or classical machine learning methods are used, and these methods cannot strengthen important features and dig deep information, resulting in insufficient tampering detection accuracy. Since the ENF in audio fluctuates randomly with time, making full use of the temporal information of ENF can improve the representation ability of features. Based on this, we propose a digital audio tampering detection method based on the deep temporal–spatial features of ENF.

3. Preliminaries

In this section, we first formally define the digital audio tampering detection task. Next, we explain the definitions related to ENF shallow temporal and spatial features and ENF deep temporal and spatial features. The mathematical notations and descriptions are shown in

Table 1. Mathematical notations and descriptions.

Notations	Descriptions
$v\left[n\right]$, $v_d\left[n\right]$, $v_{ENFC}\left[n\right]$, n	Digital audio signal, downsampled signal, ENFC signal, n is indexed
$f_d$	Downsampling frequency
${DFT}^1$, ${DFT}^k$	1- order $DFT$, k- order $DFT$
$v\left[k\right]$, $v^{\prime }\left[k\right]$	Signal after $DFT$ of $v_{ENFC}\left[n\right]$ and $v_{ENFC}^{\prime }\left[n\right]$
k, $k_{peak}$	Indexing of signal and signal peak points per frame
${\varphi }_0$, ${\varphi }_1$	0th order phase sequence, 1st order phase sequence
$f_{DF{T}^1}$	Frequency value of the first-order ENF signal
$T_{p_n\times f_n}$	Shallow temporal feature of ENF
$S_{m\times m}$	Shallow spatial feature of ENF
$floor$, $ceil$	Rounding down, rounding up
$overlap(.)$	Calculation formula of the frameshift
$F(.)$	Dilated convolution formula
$Loss$	Binary cross-entropy loss function
$ACC$, $F1\text{-}score$	Prediction accuracy, F1-score

. They are explained in more detail in the following sections.

3.1. Problem Definition

Definition 1. 

(Digital audio tampering detection task). The digital audio tampering detection task is divided into a training phase and a testing phase. In the training phase, our training audio is divided into tampered audio and untampered audio, after which these two classes of audio are trained to obtain the audio tampering detection model. In the testing phase, the test audio is fed into the audio tampering detection model to obtain two scores, i.e., the probability of belonging to the tampered and untampered categories. When $S^{*}=Score1$, the audio belongs to the untampered audio class; when $S^{*}=Score2$, the audio belongs to the tampered audio class. The flowchart of the digital audio tampering detection task is shown in Figure 1.

3.2. ENF Shallow Temporal and Spatial Feature Definition

Definition 2. 

(ENF shallow temporal features). We frame the ENF phase based on the way the ENF temporal variation is performed to obtain the ENF shallow temporal feature $T_{p_n\times f_n}$, where $p_n$ is the number of phase points contained in a frame and $f_n$ is the number of frames.

Definition 3. 

(ENF shallow spatial features). Since the ENF phase changes abruptly when audio tampering occurs, we design a framing method for the static features of the phase sequence to extract the shallow spatial features $S_{m\times m}$ of the ENF, where m is the dimension of the shallow spatial features.

3.3. ENF Deep Temporal and Spatial Feature Definition

Definition 4. 

(ENF deep temporal features). Deep temporal features are extracted from $T_{p_n\times f_n}$ by RDTCN. Deep temporal features are extracted from ENF shallow temporal features using deep learning, which requires the use of networks with sequence modeling capabilities, such as Recurrent Neural Networks (RNN) and Long Short-Term Memory (LSTM). In this paper, we use a modified TCN to extract deep temporal features that reflect the temporal information in the feature data, including the analysis of similarity information and mutation information between neighboring frames, as well as the analysis of long-term fluctuation information.

Definition 5. 

(ENF deep spatial features). Deep spatial features are extracted by CNN from $S_{m\times m}$. The deep spatial features of ENF extracted by the CNN have a larger perceptual field and richer information than the shallow features.

4. Methods

We propose a digital audio tampering detection method based on ENF deep temporal–spatial features. The method is mainly divided into two parts: ENF shallow temporal and spatial features extraction and the construction of a parallel RDTCN-CNN network model. In ENF shallow temporal and spatial feature extraction, firstly, the ENF first-order phase sequence feature ${\varphi }_1$ is extracted based on ${DFT}^1$[6]; then, it is divided into frames by adaptive frameshifting to obtain a phase feature matrix of the same size to represent the spatial features of ENF; at the same time, it is divided into frames according to the ENF phase temporal change information to represent the temporal features of the ENF. In the parallel RDTCN-CNN network model, we first extract deep temporal and spatial features using RDTCN and CNN, respectively, then use the branch attention mechanism to achieve deep temporal and spatial feature fusion, and finally complete the tampering detection by the MLP network. The framework of the digital audio tampering detection method based on ENF deep temporal–spatial feature is shown in Figure 2.

4.1. Extraction of the Shallow Temporal and Spatial Features of ENF

Because ENF has temporal and spatial information, we can extract it by different frame processing methods. The steps of extracting ENF shallow temporal and spatial features include: extracting ENF first-order phase sequence, extracting ENF shallow temporal features, and extracting ENF shallow spatial features.

4.1.1. Extraction of First-Order Phase Features

When the digital audio has been tampered with, the phase of the ENF changes abruptly at the tampered position, as shown in Figure 3. We achieve tampering detection of digital audio by representing the ENF phase features and performing consistency analysis on them.

To extract the first-order phase feature ${\varphi }_1$ of the ENF, the ENFC in digital audio is first extracted by downsampling and band-pass filtering. Then, the ENFC is subjected to $DF{T}^1$ to estimate the ENF first-order phase sequence feature ${\varphi }_1$, where $DF{T}^k$ denotes the $DFT$ transform of the kth-order derivative of the signal [6].

To extract the ENFC from the digital audio, firstly, the digital audio signal $v\left[n\right]$ to be measured is downsampled to obtain the downsampled signal $v_d\left[n\right]$, where the downsampling frequency $f_d$ is set to 1000 Hz or 1200 Hz. Secondly, we use bandpass filtering to narrowband filter the downsampled signal $v_d\left[n\right]$ to obtain the ENFC $v_{ENFC}\left[n\right]$ in the signal to be measured.

To estimate the phase information more accurately, we use $DF{T}^1$ to further extract the first-order phase sequence information ${\varphi }_1$. First, we calculate the approximate first-order derivatives of the ENF signal $v_{ENFC}\left[n\right]$ at point n. The equation to obtain the first-order derivative signal $v_{ENFC}^{{}^{\prime }}\left[n\right]$ is as follows:

$$v_{ENFC}^{\prime }\left[n\right]=f_d\left(v_{ENFC}\left[n\right]-v_{ENFC}[n-1]\right)$$

(1)

Secondly, a window is added to the first-order derivative signal $v_{ENFC}^{{}^{\prime }}\left[n\right]$. Then, the $DFT$ transform is applied to $v_{ENFC}\left[n\right]$ and $v_{ENFC}^{{}^{\prime }}\left[n\right]$ to obtain $v\left[k\right]$ and $v^{{}^{\prime }}\left[k\right]$, respectively. We estimate the frequency values based on $v^{{}^{\prime }}\left[k_{peak}\right]$ as follows:

$$f_{DF{T}^1}=\frac{1}{2\pi }\frac{DF{T}^1\left[k_{\mathrm{peak}}\right]}{DF{T}^0\left[k_{\mathrm{peak}}\right]},$$

(2)

where $DF{T}^0\left[k_{\mathrm{peak}}\right]=v\left[k_{\mathrm{peak}}\right]$ and $DF{T}^1\left[k_{\mathrm{peak}}\right]=F\left(k\right)v^{\prime }\left[k_{\mathrm{peak}}\right]$. $f\left(k\right)$ is a scale function. Finally, we use the $DF{T}^1$ method to estimate ${\varphi }_1$, which is as follows:

$${\varphi }_1=arctan\left\{\frac{tan\left(\alpha \right)\left[1-cos\left({\omega }_0\right)+sin\left({\omega }_0\right)\right]}{1-cos\left({\omega }_0\right)-tan\left(\alpha \right)sin\left({\omega }_0\right)}\right\},$$

(3)

where ${\omega }_0\approx 2\pi f_{DF{T}^1}/f_d$. There are two possible values of the solution of ${\varphi }_1$. In this paper, we use ${\varphi }_0$ as a reference and choose the value of ${\varphi }_1$ that is closest to ${\varphi }_0$ as the final solution. We perform linear interpolation of $v^{{}^{\prime }}\left[k\right]$ to obtain the value of $\alpha$. $\alpha$ is calculated as follows:

$$\alpha \approx \left(k_{DF{T}^1}-k_{\mathrm{low}}\right)\frac{{\alpha }_{\mathrm{high}}-{\alpha }_{\mathrm{low}}}{k_{\mathrm{hig}\hslash }-k_{\mathrm{low}}}+{\alpha }_{\mathrm{low}},$$

(4)

where $k_{low}=floor\left[k_{DF{T}^1}\right]$, $k_{high}=ceil\left[k_{DF{T}^1}\right]$. $floor\left[i\right]$ represents the largest integer less than i, and $ceil\left[j\right]$ represents the largest integer greater than j. $k_{DF{T}^1}=f_{DF{T}^1}{N}_{DFT}/f_d$, and $f_d$ is the downsampling frequency. The specific process is shown in Algorithm 1.

Algorithm 1 Extraction of first-order phase features ${\varphi }_1$.

Input:  ENF component: $v_{ENFC}\left[n\right]$, Down sampling frequency: $f_d$ Output:  ${\varphi }_1$ 1: Calculate the first derivative of $v_{ENFC}^{\prime }\left[n\right]$: $$v_{ENFC}^{\prime }\left[n\right]=f_d\left(v_{ENFC}\left[n\right]-v_{ENFC}[n-1]\right).$$ 2: Add Hanning windows to $v_{ENFC}\left[n\right]$ and $v_{ENFC}^{\prime }\left[n\right]$ respectively: $$v_N\left[n\right]=v_{ENFC}\left[n\right]w\left(n\right),$$ $$v_N^{\prime }\left[n\right]=v_{ENFC}^{\prime }\left[n\right]w\left(n\right).$$ 3: $DFT$ transform $v_N\left[n\right]$ and $v_N^{\prime }\left[n\right]$ to get $v\left[k\right]$ and $v^{\prime }\left[k\right]$: $$v\left[k\right]=DFT\left(v_N\left[n\right]\right),$$ $$v^{\prime }\left[k\right]=DFT\left(v_N^{\prime }\left[n\right]\right).$$ 4: By using the maximum value $k_{peak}$ of each frame signal as the integer index of $\left|v\right[k\left]\right|$ and $\left|v^{\prime }\left[k\right]\right|$, $v\left[k_{peak}\right]$ and $v^{\prime }\left[k_{peak}\right]$ can be obtained. 5: Calculating zero-order phase sequence features ${\varphi }_0$: $${\varphi }_0=arg\left[v\left(k_{\mathrm{peak}}\right)\right].$$ 6: Calculating the frequency sequence of the ENF: $$f_{DF{T}^1}=\frac{1}{2\pi }\frac{DF{T}^1\left[k_{\mathrm{peak}}\right]}{DF{T}^0\left[k_{\mathrm{peak}}\right]}.$$ 7: Calculating first-order phase sequence features ${\varphi }_1$: $${\varphi }_1=arctan\left\{\frac{tan\left(\alpha \right)\left[1-cos\left({\omega }_0\right)+sin\left({\omega }_0\right)\right]}{1-cos\left({\omega }_0\right)-tan\left(\alpha \right)sin\left({\omega }_0\right)}\right\}$$ $$\alpha \approx \left(k_{DF{T}^1}-k_{\mathrm{low}}\right)\frac{{\alpha }_{\mathrm{high}}-{\alpha }_{\mathrm{low}}}{k_{\mathrm{hig}\hslash }-k_{\mathrm{low}}}+{\alpha }_{\mathrm{low}}$$ 8: return ${\varphi }_1$

4.1.2. Extraction of ENF Shallow Temporal Features

Various information and phase sequences of ENF change with its non-periodic fluctuations. In this paper, we propose a new framing algorithm based on the way of ENF timing variation to obtain the ENF shallow temporal feature $T_{p_n\times f_n}$, where $p_n$ is the number of phase points contained in a frame, which is an artificially set value. Since the setting of $p_n$ affects the amount of information contained in the ENF phase timing over a period of time, we will experiment with $p_n$ as a variable for the ENF temporal representation in the subsequent part of this paper. In addition, digital audio is often unequal in length, corresponding to its ENF first-order phase sequence. To reduce the loss of ENF temporal information, we calculate the number of frames from the longest audio. Based on the set frame length $p_n$, the number of frames $f_n$ can be calculated as follows:

$$f_n=\frac{M_{max}}{p_n},$$

(5)

where $M_{max}$ is the maximum value of the ENF phase sequence. The frameshift $overlap$ is calculated as follows:

$$overlap=p_n-floor\left[\frac{length\left({\varphi }_1\right)}{f_n}\right],$$

(6)

where the $floor$ is rounded down. Finally, the shallow temporal feature $T_{p_n\times f_n}$ is obtained by traversing the phase information of ENF by frameshift $overlap$. The specific process is shown in Algorithm 2.

Algorithm 2 Temporal feature frame processing of ENF.

Input:  Phase sequence features: ${\varphi }_1$, the number of phase points contained in a frame: $p_n$ Output:  Shallow temporal features of ENF: $T_{p_n\times f_n}$ 1: Calculate the length of the phase sequence $M_{max}$. 2: Calculate the number of frames: $f_n=\frac{M_{max}}{p_n}$. 3: for Phase sequence features of all audio do 4:    Calculate the frameshift $overlap$: $$overlap=p_n-floor\left[\frac{length\left({\varphi }_1\right)}{f_n}\right].$$ 5:    Split the frame. 6:    Reshape into feature matrix $T_{p_n\times f_n}$. 7: end for 8: return $T_{p_n\times f_n}$

4.1.3. Extraction of ENF Shallow Spatial Features

When digital audio is tampered with, the ENF phase features cause abrupt changes [17]. To represent the static mutation information in the phase sequence, we designed a framing method for the static features of the phase sequence to extract the shallow spatial feature $S_{m\times m}$ of the ENF. The extracted phase sequence features ${\varphi }_1$ are of different lengths due to the unequal duration of each sample in the dataset. To reduce the information loss in the feature extraction process, we calculate the frame length m by the longest phase $len\left({\varphi }_{DF{T}^1}\right))$ in the audio, which is as follows:

$$m=ceil\left(\sqrt{X}\right),$$

(7)

where $X=ceil\left(\sqrt{len\left({\varphi }_{DF{T}^1}\right)}\right)$, and $ceil$ is rounded upward. Then, the frameshift and subframe are calculated, and the frameshift $overlap$ is as follows:

$$overlap=m-ceil\left(\frac{X-m}{m-1}\right)$$

(8)

The shallow spatial feature matrix $S_{m\times m}$ is obtained by giving an adaptive frameshift of unequal audio phase sequences to facilitate automatic learning of the CNN. The specific process is shown in Algorithm 3.

Algorithm 3 Spatial feature frame processing of ENF.

Input:  Phase sequence features: ${\varphi }_1$ Output:  Shallow spatial features of ENF: $S_{m\times m}$ 1: Calculate the longest phase $len\left({\varphi }_{DF{T}^1}\right))$. 2: Calculate the number of frames m: $$m=ceil\left(\sqrt{X}\right),\mathrm{where}X=ceil\left(\sqrt{len\left({\varphi }_{DF{T}^1}\right)}\right).$$ 3: for Phase sequence features of all audio do 4:    Calculate the frameshift $overlap$: $$overlap=m-ceil\left(\frac{X-m}{m-1}\right).$$ 5:    Split the frame. 6:    Reshape into feature matrix $S_{m\times m}$. 7: end for 8: return $S_{m\times m}$

4.2. Deep Feature Representation Learning Based on RDTCN-CNN Temporal–Spatial Feature Fusion

Based on $T_{p_n\times f_n}$ and $S_{m\times m}$ extracted in Section 4.1, we design a parallel RDTCN-CNN model to implement digital audio tampering passive detection, as shown in step 2 of Figure 2. The parallel RDTCN-CNN model is divided into three main stages: extraction of deep temporal and spatial features, the fusion of temporal and spatial features, and classification decision.

In the extraction stage of deep temporal and spatial features, we input the shallow temporal features $T_{p_n\times f_n}$ into the RDTCN for representation learning to obtain the deep temporal features, while the shallow spatial features $S_{m\times m}$ are input into the CNN for representation learning to obtain the deep spatial features. In the fusion stage of deep temporal and spatial features, we use the branch attention mechanism to fuse the deep temporal and spatial features of different branches to obtain the temporal–spatial features with stronger representation capability. In the classification decision stage, we input the temporal–spatial features to the MLP network to determine whether the audio has been tampered with.

4.2.1. Deep Temporal Feature Extraction Based on RDTCN Network

As the network deepens, the ordinary TCN network comes with residual layers and each convolutional layer has different layer-size perceptual fields. However, the deep temporal feature extraction network based on the ordinary TCN network neglects to make full use of the information of different receptive fields in each convolutional layer. This means that the common TCN network has a common residual structure, which causes the intermediate convolutional layers to not directly transfer information to the subsequent layers, so the memory block does not fully utilize the information of all the convolutional layers inside it. In order to improve the disadvantage of underutilized ordinary TCN memory blocks, we construct an RDTCN network using residual dense blocks and then extract the ENF deep temporal features. The RDTCN further improves information utilization compared with the ordinary TCN network. The RDTCN network structure is shown in Figure 4.

To address the shortcomings of ordinary residual networks, such as the inability to utilize feature information between each layer in the ENF deep temporal feature extraction task, a residual dense network is proposed instead of the ordinary residual blocks. The residual dense network is composed of a residual network and a dense network, with the residual network extracting global features and the dense convolutional layer extracting local features. In the dense connection, each layer of information fusion combines the inputs of all previous layers and uses cascading to propagate the features of the current layer to all subsequent layers, which is more effective in mitigating gradient disappearance and enhancing the propagation of all features, making the probability of feature reusability greater. The residual dense network combines the characteristics of the residual network and dense network to form a continuous memory mechanism by fusing global features and local features, which is implemented by linking the features extracted from the previous residual dense block to all layers of the current residual dense block. The residual dense block supports continuous memory, and after extracting multiple layers of local dense features, it will further fuse the global features and then retain the layered features in a global manner adaptively, thus producing implicit deep supervision.

The RDTCN mainly consists of causal convolution, dilated convolution, and residual dense modules. In causal convolution, a sequence $T=\left(T_1,T_2,\dots ,T_t\right)$ is input, and a sequence $Y=\left(Y_1,Y_2,\dots ,Y_t\right)$ is output after the operation of the model. The output at time t is only convolved with the historical elements before time t, thus ensuring that the future input will not affect the past input data prediction.

By adding holes to the standard convolution to expand the field of perception, the output data can contain a larger range of information without losing data in the pooling layer. The dilated convolution formula is as follows:

$$F\left(s\right)=\left(T{\ast }_{d}k\right)\left(s\right)=\sum _{t=0}^{c-1}k\left(i\right)·T_{s-d·i},$$

(9)

where the input sequence $T=\left(T_1,T_2,\dots ,T_t\right)$, filter $k=\left(k_1,k_2,\dots ,k_c\right)$, d is the expansion factor, c is the filter size, and $s-d·i$ denotes the past direction. The value of d relates to the size of the perceptual field, and d grows exponentially by 2.

In the residual dense block, each residual dense block is a combination of Conv -BN-ReLU, and the residual block is as follows:

$${\overline{Y}}_R^l=H_l\left(\left[x_{l-1},x_{l-1}^0,x_{l-1}^1,\cdots ,x_{l-1}^{N-1}\right]\right),$$

(10)

where $x_{l-1}$ is the input of the lth residual dense block, $x_l$ is the output of the lth residual dense block, and $H_l$ denotes the last convolutional layer that performs the splicing operation on the output of the feature from all the convolutional layers in the brackets. The structure of the residual dense network is shown in the residual dense block in Figure 4, where $Conv$ means convolution, $BN$ means batch normalization, $ReLU$ means activation function, and the convolution kernel size is $3\ast 3$ with a step size of 1.

4.2.2. Deep Spatial Feature Extraction Based on CNN Network

In this section, in order to extract the deep spatial features of ENF quickly and efficiently, we design a CNN to implement the extraction of deep spatial features, as shown in Figure 5.

CNN is a multilayer feedforward neural network that has been shown to have outstanding performance in extracting spatial features. The sparse connectivity and weight-sharing nature of the CNN greatly reduce the number of model parameters. The sparse connectivity and weight-sharing nature of the CNN allow it to learn audio features with less computational effort, with stable results, and without additional feature engineering requirements on the data.

The CNN consists of a convolutional layer and a pooling layer. The convolutional layer further extracts the deep spatial features by convolving the input shallow spatial features. The input to the deep space feature extraction module of CNN is $S_{m\times m}$, which has dimensions (45, 45). The deep spatial feature extraction module consists of three convolutional layers, three pooling layers, and two fully connected layers. The number of convolutional kernels is 16, 32, 64, and the size of the kernels is (3, 3). $3\ast 3$ convolutional kernels can increase the number of network layers compared to larger convolutional kernels, which also increases the nonlinear expression capability of the network and reduces the network parameters. After each convolutional layer, a MaxPooling layer of size (2, 2) is set. The high-level feature map obtained after pooling not only reduces the dimensionality and number of parameters of the original feature map, but also avoids problems such as overfitting. After the convolutional pooling of multiple layers, the Flatten layer is connected to transform the multidimensional data into one-dimensional data to realize the transition from the convolutional layer to the fully connected layer. The number of nodes in the first fully connected layer is 1024, and the number of nodes in the second fully connected layer is 256. The convolutional layer cooperates with the pooling layer to extract local features of spatial features, then connects with the fully connected layer to extract the global features and finally obtains deep spatial features.

4.2.3. Deep Temporal and Spatial Feature Fusion Based on Branch Attention Mechanism

Because deep temporal and spatial features are extracted from different branches of the parallel RDTCN-CNN, we choose the branch attention mechanism for deep temporal and spatial feature fusion to achieve the fusion of temporal and spatial features. The branch attention mechanism can achieve linear fusion of deep temporal features with deep spatial features when performing the feature fusion task. The equation of the branch attention mechanism is as follows:

$$\left\{\begin{array}{c}{R}_C=concat\left(D{T}_{p_n\times f_n},D{S}_{m\times m}\right)\hfill \\ W_b=F_b\left(R_C\right)\hfill \\ S_T=R_C{W}_b\hfill \end{array}\right.,$$

(11)

where $D{T}_{p_n\times f_n}$ and $D{S}_{m\times m}$ are, respectively, the deep temporal and spatial features, $concat(.)$ is the join operation, $R_C$ is the spliced feature and $F_b(.)$ is the convolutional pooling operation in the branch attention mechanism, and $S_T$ reflects the deep temporal–spatial features after fusion.

In addition, this mechanism assigns weights to different types of features through network learning, so that important features have more prominent representational power in model training. The branch attention fusion mechanism includes weight learning and dot product assignment. In weight learning, we learn useful information in deep temporal and spatial features by convolutional operations. In the dot product assignment, we dot product the learned weights with the spliced deep temporal–spatial and spatial features to adaptively assign the weights and obtain the temporal–spatial features with stronger representational ability, as shown in Figure 6.

4.2.4. Classification Network Design

In this section, we use the MLP classification network to determine whether it has been tampered with. The network consists of four fully connected layers (the number of neurons is 256, 128, 32, and the activation function is $LeakyReLU$), a Dropout layer (dropout rate = 0.2), and a $softmax$ layer, and the specific network structure is shown in Figure 2. $softmax$ layer can output the binary classification results for determining whether the audio has been tampered with or not, and the formula is as follows:

$${\widehat{y}}^{\left(j\right)}=softmax\left(I^{\left(j\right)}W+b\right),$$

(12)

where ${\widehat{y}}^{\left(j\right)}$ is the predicted value of the output, $I^{\left(j\right)}$ numbers the output features of the previous layer, W is the weight, and b is the bias.

Our loss function adopts binary cross entropy. This is a commonly used loss function in binary classification problems, and its expression is as follows:

$$Loss=-\frac{1}{N}\sum _{i=1}^N{y}_i·log\left(p\left(y_i\right)\right)+\left(1-y_i\right)·log\left(1-p\left(y_i\right)\right),$$

(13)

where N is the number of audio data, y corresponds to the label value of each voice, and $p\left(y\right)$ is the probability that the output belongs to label y. $Loss$ is the value of the binary cross entropy loss function, which is used to judge the performance of our model.

5. Experimental Results and Analysis

To validate the effectiveness of the proposed method, we conducted a large number of experiments on different audio datasets. To validate the effectiveness of each independent part of the framework, we followed the idea of the controlled variable method to conduct an ablation study to observe the core modules and key hyperparameters of our method.

5.1. Dataset

In this paper, Carioca [25,47], New Spanish [52], and ENF_Audio are used as experimental data. The detailed statistics of these datasets are shown in

Table 2. Dataset information.

The Dataset	Carioca	New Spanish	ENF_Audio
Edited audio	250	502	752
Original audio	250	251	501
Total audio	500	753	1253
Audio duration	9∼35 s	16∼35 s	9∼35 s
The training set	350	527	877
The validation set	50	75	125
The test set	100	151	251

.

Carioca consists of Carioca 1 [25] and Carioca 2 [47], a database of speech signals from telephone recordings on the public switched telephone network (PSTN), both with an ENF of 60 Hz. Cariocacontains 500 speech signals. The audio in Carioca 1 was all sampled at 44.1 kHz, with a word length of 16 bits, a bit rate of 705 kbps, a single channel, a file format of WAV, and a signal-to-noise ratio (SNR) ranging from 16 dB to 30 dB (22.3 dB on average). Carioca 1 has a total of 200 audio signals, and the audio signals range from 19 to 35 s in duration. There were 100 of these original audio signals, including 50 male and 50 female. For each gender, 25 signals were copied and 25 signals were cut, resulting in another 100 tampered audios, where each signal was copied or cut only once. Carioca 2 had a sampling frequency of 11,050 Hz, a word length of 16, a bit rate of 176 kbps, a single channel, a file format of WAV, and a signal-to-noise ratio of 10.1 dB to 30 dB (the average Carioca 2 file contains a total of 300 signals with durations of 9 to 25 seconds). There were 150 raw audios, including 75 male and 75 female. An additional 150 tampered signals were obtained by performing a delete or insert operation on all 150 original audios. The processing was carried out in the same way as in Carioca 1.

New Spanish [52] comes from two public Spanish databases, AHUMADA and GAUDI, containing stable 50 Hz ENFCs, which also contain 753 speech signals. All signals have a sampling rate of 8 kHz, a word length of 16, a bit rate of 128 kbps, a single channel, a file format of WAV, and an average signal-to-noise ratio of 35 dB. There were 251 original audio files. The 502 tampered audio files were obtained by deletion and insertion operations.

ENF_Audio is our database, consisting of a random mix of Carioca and New Spanish, with a total of 1253 audio signals.

5.2. Evaluation Metrics

Tampering detection of digital audio is a complex task involving multiple domains of security. To verify the reliability of the model, we cite feedback prediction as the main evaluation task. In the experiments of this paper, we use prediction accuracy (ACC) and F1-score (F1-score) as evaluation metrics.

$ACC$ is the error between the prediction result and the actual feedback, demonstrating the most significant value and progress of the model. The higher the Accuracy, the higher the precision. The formula for prediction accuracy is:

$$ACC=\frac{TP+TN}{TP+FN+FP+TN},$$

(14)

where $TP$ is predicted to be positive and actually positive, $TN$ is predicted to be negative and actually negative, $FP$ is predicted to be positive and actually negative, and $FN$ is predicted to be negative and actually positive. The ACC is calculated using the evaluate function in the Keras library. (The API address for Evaluate is at https://keras.io/api/models/model_training_apis/ (accessed on 10 April 2023))

$F1\text{-}score$ is the result of combining the output of Precision (P) and Recall (R), and the $F1\text{-}score$ ranges from 0 to 1, with 1 representing the best output of the model and 0 representing the worst output of the model. The formula for the $F1\text{-}score$ is as follows:

$$F1\text{-}score=2\frac{P\ast R}{P+R},$$

(15)

where P represents the precision rate, which is given by the formula: $P=\frac{TP}{TP+FP}$, and R represents the recall rate, which is given by the formula: $R=\frac{TP}{TP+FN}$. The F1-score is calculated using the metrics function in the Scikit-learn library. (The API address for Metrics is at https://scikit-learn.org/stable/modules/generated/sklearn.metrics.f1_score.html (accessed on 10 April 2023))

5.3. Baselines

To evaluate the performance of our model, we need other baselines for comparison. The details of the baselines are as follows: $DFT1-SVM$[25] and $ES-SVM$[26] are two classical tamper detection methods. Therefore, we use them as baselines. In addition, in our previous work, $PF-SVM$[6] is a new method proposed based on the advantages of the above two methods [25,26], so we also use it as a baseline. $X-BiLSTM$ [5] is a new audio tampering detection method from the perspective of temporal feature representation learning, so we also use it as a baseline.

$DFT1-SVM$ [25] estimated the phase of ENFC by discrete Fourier variation and used SVM for classification.

$ES-SVM$ [26] used ESPRIT to estimate the characteristics of the phase peak and used SVM for classification.

$PF-SVM$[6] used ${DFT}^0$, ${DFT}^1$, and Hilbert transform to extract the phase and instantaneous frequency of ENFC and used an optimized SVM for the classification of tamper detection.

$X-BiLSTM$[5] divides the extracted phase features into frames, and each frame is represented as the information of ENF phase change over a period of time, and then the BI-LSTM network is used to obtain the difference information between real audio and tampered audio, and finally, a DNN classifier is used for classification.

5.4. Experimental Settings

Hardware Setting: In this paper, experiments were conducted on a professional computer with an NVIDIA TITAN RTX high-performance graphics card. The CPU is an I7-9700, the GPU is an RTX TITAN X, and the running memory is 64 GB.

Software Setting: Audio framing, windowing, filtering, and parameter extraction were performed on MATLAB R2020a. Our model simulations were performed on Python 3.6, Tensorflow 1.15, Keras 2.1.5, Numpy 1.19, Pandas 0.25, etc.

Framework Setting: The shallow spatial feature $S_{m\times m}$ has an m of 46, the shallow temporal feature $T_{p_n\times f_n}$ has a $p_n$ of 85 and $f_n$ of 25, the loss function is Binary_Crossentropy, the optimizer is Adam, the training epoch is 300, the batch size is 64, and the initial learning rate is 0.001. The training set, validation set, and test set are divided into 7:1:2.

5.5. Results and Discussion

To verify the effectiveness of the method in this paper, we demonstrate the effectiveness of the features and classification models in this section and the superiority of the method through four sets of experiments. The experiments are designed as follows: (1) comparison with baseline methods; (2) verifying the effectiveness of RDTCN temporal feature extraction network; (3) verifying the effect of frame length setting on the shallow temporal features of ENF; (4) verifying the effectiveness of the branch attention mechanism.

5.5.1. Comparison with Baseline Methods

In order to verify the effectiveness of our proposed digital audio tampering detection method, we compare the proposed method RDTCN-CNN with the method of baseline [5,6,25,26] in the public Carioca, New Spanish, and ENF_Audio datasets, and the experimental results are shown in

Table 3. Comparison with baseline.

Method	Carioca		New Spanish		ENF_Audio
	ACC (%)	F1-Score (%)	ACC (%)	F1-Score (%)	ACC (%)	F1-Score (%)
DFT1-SVM [25]	89.90	90.22	88.86	86.84	90.51	90.55
ES-SVM [26]	90.88	90.62	90.62	88.26	93.52	93.44
PF-SVM [6]	93.05	92.86	90.22	87.56	92.60	92.82
X-BiLSTM [5]	97.03	97.22	92.14	90.62	97.22	97.02
RDTCN-CNN	97.96	97.54	95.60	94.50	98.02	97.88

.

From the results, it can be seen that the $ACC$ and $F1\text{-}score$ of X-BiLSTM and the method in this paper are much higher than the other three methods for the same dataset [6,25,26]. Both X-BiLSTM [5] and RDTCN-CNN use deep-learning methods, which have better performance in extracting key information in features compared to classical machine learning. For both methods using deep learning, the method in this paper uses a parallel RDTCN-CNN network to extract both spatial and temporal information of ENF. In comparison, the X-BiLSTM only utilizes temporal information in the ENF phase and ignores information on ENF phase space, and is 1.20% higher than X-BiLSTM in terms of accuracy.

As can be seen in Figure 7, the $ACC$ and $F1-score$ of RDTCN-CNN in different datasets are higher than the other four baseline methods, which verifies the effectiveness of this method.

5.5.2. Verifying the Effectiveness of the RDTCN Temporal Feature Extraction Network

To verify the effectiveness of RDTCN networks, a set of comparison experiments based on ordinary TCN temporal feature extraction and RDTCN temporal feature based on shallow spatial feature extraction networks are conducted in this section, both of which adopt the structure of CNN.

In the RDTCN temporal feature extraction network, the original residual blocks of the TCN network are replaced with residual dense blocks. From the experimental results in

Table 4. Comparison of RDTCN and ordinary TCN.

Method	Carioca		New Spanish		ENF_Audio
	ACC (%)	F1-Score (%)	ACC (%)	F1-Score (%)	ACC (%)	F1-Score (%)
Ordinary TCN	96.58	96.22	93.56	91.88	97.42	97.46
RDTCN	97.96	97.54	95.60	94.50	98.02	97.88

, it can be seen that the ordinary TCN temporal feature extraction network based on the RDTCN achieves 97.42%, while the TCN network with the addition of residual dense blocks achieves 98.02%, which is an accuracy improvement of 0.6% to 98.02% compared to the ordinary TCN temporal feature extraction network. In addition, the RDTCN temporal feature extraction network also improves the $F1-score$ by 0.42% compared to the normal TCN temporal feature extraction network. Different evaluation metrics are sufficient to prove that the RDTCN temporal feature extraction network is effective and that the features in the layered convolution in the TCN network are effective for the digital audio tampering detection task.

5.5.3. Verifying the Effect of Frame Length Setting on the Shallow Temporal Features of ENF

The digital audio signal is a time-varying signal, and various information of audio and parameters characterizing its basic features change with time but remain basically unchanged in a short period of time, so the audio signal has long-term fluctuation and short-term stability. Every 0.17 s contains 10 phase points, and the phase number $p_n$ determines the amount of information of ENF phase change in a short time. A suitable frame length setting can better represent the fluctuation of temporal information and is more conducive to the extraction of temporal information by the RDTCN network. In order to analyze the effect of frame length on features, we conducted comparison experiments on the frame length settings of ENF shallow temporal features on Carioca, New Spanish, and ENF_Audio datasets, and nine experimental groups were set in the frame length range of 0.255 s∼1.615 s, and the accuracy variation curves with frame length are shown in Figure 8.

From the experimental results, we can see that the detection accuracy is high when the frame length is 0.595 s and the number of phase points $p_n$ per unit frame is 35, the accuracy tends to decrease after the frame length is 1.445 s, and the number of phase points $p_n$ per unit frame is 85. Since the highest accuracy is achieved when the frame length is 1.445 s and the number of phase points per unit frame $p_n$ is 85, we choose the frame length of 1.445 s, the number of phase points per unit frame $p_n$ of 85, and the number of frames of 25 as the parameters of the ENF shallow temporal feature $T_{p_n\times f_n}$.

5.5.4. Verifying the Effectiveness of the Branch Attention Mechanism

To verify the effect of branch attention mechanism on parallel RDTCN-CNN, we designed a comparison experiment between the feature fusion network based on the splicing structure and the feature fusion network based on the branch attention mechanism. The former uses the splicing layer to replace the attention mechanism module to achieve the fusion of deep temporal and spatial features, and the other network parameters remain unchanged.

The experimental results are shown in

Table 5. A comparative experiment of branch attention mechanisms.

Method	Carioca		New Spanish		ENF_Audio
	ACC (%)	F1-Score (%)	ACC (%)	F1-Score (%)	ACC (%)	F1-Score (%)
Splice Fusion	96.02	96.42	94.42	92.82	97.20	97.22
Branch	97.96	97.54	95.60	94.50	98.02	97.88

. With our parallel RDTCN-CNN network, based on the attention mechanism, the Carioca, New Spanish, and ENF_Audio datasets all show significant improvement in accuracy compared to the parallel RDTCN -CNN network based on the splicing structure. This proves that the branch attention mechanism is effective for different types of feature weight distributions. Through adaptive learning, the attention mechanism can give more weight to the features that are useful for classification and suppress the invalid features, thus improving detection accuracy.

6. Conclusions

In this paper, we propose a digital audio tampering detection method based on ENF deep temporal–spatial features. Structurally, our method includes the extraction of shallow temporal and spatial features of ENF and the construction of parallel RDTCN-CNN network models. For the extraction of shallow temporal and spatial features, the phase sequence of ENF is first extracted by using the high-precision discrete Fourier analysis method. Secondly, for the information on different dimensions of ENF, shallow temporal and spatial features are extracted using different frame processing methods. In the construction of a parallel RDTCN-CNN network model, we use RDTCN, which is good at processing temporal signals, to further extract deep temporal features, and use CNN network, which is good at extracting spatial features, to further extract deep spatial features, and realize the fusion of temporal–spatial features by the branch attention mechanism. Finally, the MLP network is used to determine whether the digital audio has been tampered with or not. The experimental results show that our proposed method has a high accuracy and F1-score in the Carioca, New Spanish, and ENF_Audio databases, outperforming the four baseline methods.

The method in this paper represents only the phase features of the ENF, while the harmonic signal, spectrogram, and many other features of the ENF should be equally valued. Moreover, the research in this paper focuses on whether the audio has been tampered with and only addresses the basic issues. Further analysis is still needed to determine the deeper location of tampering and the specific type of tampering. Therefore, in future work, deeper mining of more features in the ENF and further analysis of tampering locations and tampering types are necessary. In addition, we need to further automate the filtering and feature extraction of the original audio waveform through the network model and implement end-to-end tasks in the model to make it applicable to more complex scenarios to facilitate applications.