An aac security steganography algorithm and system based on masking effect

Abstract

The application discloses an AAC security steganography algorithm and system based on a masking effect. The human ear hearing masking effect can cause some audio signals with low intensity to be masked by signals with high intensity, and there is space for hiding secret information. Therefore, the application analyzes a quantization process of MDCT coefficients in AAC coding, records the masked audio signals as modifiable positions, and combines an STC adaptive steganography framework to realize embedding of secret information. Experimental results show that the algorithm can achieve a maximum embedding capacity of 13.61 kbps, can guarantee auditory concealment of speech, and has good security.

Description

Technical Field

[0001] This invention relates to the field of digital media processing technology, and particularly to a steganography algorithm for AAC audio, enabling secure transmission of confidential information in the field of multimedia information content security. Specifically, it relates to an AAC secure steganography algorithm and system based on the masking effect. Background Technology

[0002] In recent years, with the rapid development of information technology, the scope of internet use has expanded significantly. Multimedia communication, with its ease of use and convenience, has become an indispensable communication method in people's daily work and life. Audio information transmission is frequently involved in multimedia communication. Before transmitting audio over a network, the original audio information is generally compressed and encoded to address the problem of excessive network bandwidth consumption, while also ensuring audio quality during compression. AAC compressed audio coding, short for Advanced Audio Coding, is an advanced audio coding technology jointly developed by Dolby Laboratories and other companies in 1997 based on MPEG-2. The AAC encoding process involves many processing modules. Among these, the psychoacoustic model, filter bank module, scaling factor module, quantizer, noise-free coding module, and the final bitrate output module are mandatory. Other modules can be selected based on hardware devices and application scenarios. Specifically, when using AAC encoding, users can choose different specifications based on different hardware levels and sound quality requirements. The main specifications include the main frame, LC (low complexity frame), and SSR (scalable sampling rate frame). The main frame encoding uses all encoding modules except the gain control module to ensure the best audio quality, but it has high hardware requirements. LC (Low Complexity) encoding is simpler, removing the gain control and predictor modules during encoding, thus improving encoding efficiency and making it suitable for situations with lower-end hardware. SSR (Scalable Sample Rate) encoding is more suitable for situations with unstable network bandwidth. Compared to previous encoding techniques, AAC encoding can achieve higher sound quality at the same compression rate, while also improving decoding efficiency and supporting more bit rates and sampling rates. It is currently widely used in online streaming media.

[0003] AAC encoding involves three embeddable fields: scaling factor, QMDCT coefficients, and Huffman coding. Currently, most AAC steganography algorithms targeting the QMDCT coefficient field utilize LSB bits or sign bits for embedding, making them non-adaptive. Therefore, this paper analyzes and studies the encoding principle of AAC encoders, combining audio masking effects and the STC adaptive steganography framework. By selecting the QMDCT coefficient field, an AAC-based steganography algorithm is designed and implemented, enabling the embedding of secret information into AAC audio—a novel information hiding method. Summary of the Invention

[0004] The above-mentioned technical problems of the present invention are mainly solved by the following technical solutions:

[0005] A secure steganography algorithm based on the masking effect.

[0006] The original audio is encoded to obtain the QMDCT coefficient sequence Q, and then converted to obtain the binary sequence Q′.

[0007] The original audio is transformed to obtain the masked frequency bands in each frame of audio, and these are used as modifiable positions. Based on the obtained modifiable positions, the distortion cost function is used to calculate the single-point distortion cost of the QMDCT coefficients of each frequency band, and the distortion cost sequence Cost is obtained.

[0008] By inputting the secret message sequence m′, the QMDCT coefficient sequence Q′, and the distortion cost sequence Cost into the STC framework, the embedding method with the minimum overall distortion is obtained, and the final secret-containing QMDCT carrier sequence P′ is output.

[0009] In the aforementioned AAC secure steganography algorithm based on the masking effect, the original WAV audio is input into an AAC encoder for encoding. The MDCT transform is then performed, and the complete QMDCT coefficient sequence Q is obtained in the quantization module: {Q1, Q2, ..., Q...}. n}

[0010] In the aforementioned AAC secure steganography algorithm based on the masking effect,

[0011] The spectrum obtained by performing a short-time Fourier transform on the original audio.

[0012] The STFT coefficients of each frame of audio are calculated, converted into sound pressure level, and the masking coefficient of each peak signal on all other signals is calculated.

[0013] The masking threshold for all frequency bands is calculated using masking coefficients;

[0014] The masking threshold and sound pressure level of each frequency band are compared. If the signal strength of a frequency band is lower than the masking threshold, then the frequency band is the masked frequency band.

[0015] Masking refers to a frequency band whose intensity (sound pressure level, measured in dB, describes the intensity of a sound signal) is so high that the human ear cannot hear sounds in the surrounding frequency bands.

[0016] Therefore, for each frequency band, there are two values: sound pressure level (i.e., its inherent signal strength, which is present) and masking threshold (i.e., the masking strength exerted on it by surrounding frequency bands, which is calculated). If the sound pressure level of a frequency band is less than its masking threshold, then it can be considered to be masked.

[0017] In the aforementioned AAC secure steganography algorithm based on the masking effect, the masking threshold within each frequency band is the sum of the masking coefficients of all peak signals.

[0018] In the aforementioned AAC secure steganography algorithm based on the masking effect, the distortion cost sequence Cost is: {Cost1, Cost2, ..., Cost...} n The following formula is used to calculate}

[0019]

[0020] In the aforementioned AAC secure steganography algorithm based on the masking effect, the secret message m is converted into a binary bit stream sequence, and then scrambled using the key to obtain m′.

[0021] In the aforementioned AAC secure steganography algorithm based on the masking effect,

[0022] The vector sequence P′ is converted into the QMDCT coefficient sequence P through an inverse mapping operation;

[0023] The QMDCT coefficients are replaced in the quantization module of the second encoding process;

[0024] Based on the replaced QMDCT coefficients, a dense AAC audio is encoded.

[0025] The conversion between binary bitstreams and QMDCT coefficients is a routine process in AAC encoding and decoding.

[0026] Obtaining AAC audio from QMDCT coefficients is also a standard procedure in the AAC encoding process, as described in the manual. Figure 1 The (AAC encoding process) describes the modules that should be included in a complete AAC encoding framework. This method only edits the QMDCT coefficients (i.e., Figure 1The "quantizer" step in the process remains completely unchanged, while the other steps are exactly the same.

[0027] A system, including

[0028] The first module is configured to encode the raw audio, obtain the QMDCT coefficient sequence Q, and convert it to obtain the binary sequence Q′.

[0029] The second module is configured to transform the original audio to obtain the masked frequency bands in each frame of audio, and use them as modifiable positions. Based on the obtained modifiable positions, the distortion cost function is used to calculate the single-point distortion cost of the QMDCT coefficients of each frequency band to obtain the distortion cost sequence Cost.

[0030] The third module is configured to input the secret message sequence m′, the QMDCT coefficient sequence Q′, and the distortion cost sequence Cost into the STC framework to obtain the embedding method with the minimum overall distortion, and output the final secret-containing QMDCT carrier sequence P′.

[0031] This invention analyzes the quantization process of MDCT coefficients in AAC encoding, records the masked audio signal as a modifiable position, and combines it with the STC adaptive steganography framework to embed secret information. Experimental results show that the algorithm can achieve a maximum embedding capacity of 13.61 kbps while ensuring auditory concealment of speech and possessing good security. Attached Figure Description

[0032] Appendix Figure 1 This is the AAC encoding flowchart;

[0033] Appendix Figure 2 It is a frequency domain masking effect diagram;

[0034] Appendix Figure 3 These are three types of audio curves;

[0035] Appendix Figure 4 This is a flowchart for obtaining the masked frequency band;

[0036] Appendix Figure 5 It is an audio spectrogram.

[0037] Appendix Figure 6 These are the modifiable positions in the original audio under different thresholds H: (a) H=16, (b) H=10, (c) H=5, and (d) H=0.

[0038] Appendix Figure 7 These are audio QMDCT coefficient distribution plots. (a) is the frequency histogram of the first-order statistical distribution of the QMDCT coefficients, and (b) is the frequency histogram of the first-order statistical distribution of the QMDCT coefficients.

[0039] Appendix Figure 8 These are comparison charts of the statistical characteristics of QMDCT coefficients before and after steganography, where (a) is a comparison chart of the first-order statistical distribution of QMDCT coefficient frequencies, and (b) is a comparison chart of the first-order statistical distribution of frequencies.

[0040] Appendix Figure 9 This is a diagram of the steganography algorithm framework (where Modified Discrete Cosine Transform (MDCT) is a normal step in the AAC encoding and decoding process, and QMDCT coefficients are values ​​obtained by quantizing the results of the above MDCT transform). Detailed Implementation

[0041] The technical solution of the present invention will be further described in detail below through embodiments and in conjunction with the accompanying drawings.

[0042] Example:

[0043] This invention addresses the problem that there are relatively few AAC steganography methods for the QMDCT coefficient domain, and most of them are non-adaptive steganography. It implements an AAC secure steganography algorithm based on the masking effect for QMDCT coefficients.

[0044] The following is a detailed introduction.

[0045] First, let's discuss the main core principles of this invention.

[0046] 1. Steganography design.

[0047] 1.1 Analysis of modifiable QMDCT coefficients in AAC encoding.

[0048] 1.1.1 Analysis of the masking effect in AAC encoding.

[0049] The masking effect of audio signals refers to the phenomenon where, when multiple sound signals exist in an environment, we may only hear some of the louder signals, while the rest are masked. In this process, the audible sounds are called masking sounds, and the inaudible sounds are called masked sounds. In other words, the masking effect alters how the human ear perceives sound signals in the environment. When both masking and masked sounds occur simultaneously, the masking sound, due to its higher signal strength, can mask nearby sounds with lower signal strength, such as... Figure 2 As shown.

[0050] exist Figure 2In the diagram, the taller gray bar in the middle represents the masking signal, while the two shorter white bars on either side represent the masked signals. The dashed line represents the absolute hearing threshold curve, and the solid line represents the change in the threshold curve when the masking signal affects human auditory perception after the frequency domain masking effect occurs. The absolute hearing threshold curve refers to the minimum signal intensity required for the human ear to hear sound signals of different frequencies in an absolutely quiet environment. If the intensity of the sound signal is less than the absolute hearing threshold, the human ear will not be able to perceive the sound signal.

[0051] To investigate the frequency domain masking effect in AAC audio, we analyzed the spectral data of an audio segment, such as... Figure 3 As shown in the diagram. The horizontal axis represents the frequency of the sound signal, and the vertical axis represents the sound pressure level (SPL), which is the intensity of the sound signal. The orange curve shows the absolute threshold curve for human hearing, representing the minimum sound intensity at which the human ear can detect a sound signal of that frequency in a quiet environment. This curve can be approximately fitted using Equation 1-1. The green curve represents the spectral curve of the PCM audio data before encoding, representing the intensity of the sound signal at each frequency in the PCM data. The red curve represents the final threshold curve in AAC encoding.

[0052]

[0053] As shown in the figure, the final threshold curve in AAC encoding differs significantly from the absolute hearing threshold curve of the human ear in the low-frequency range, while it approximates the human ear in the high-frequency range. This is because the main information of audio is concentrated within the human ear's auditory range, i.e., the low-frequency domain. Therefore, the masking effect also mainly occurs in the low-frequency domain. Furthermore, due to the narrow bandwidth of the masking effect, peak signals in the low-frequency domain have little impact on the masking threshold of the high-frequency audio signal. Therefore, we can conclude that during AAC encoding, the auditory threshold of the audio signal does indeed change due to the presence of peak signals and the frequency domain masking effect.

[0054] 1.1.2 QMDCT coefficients can be modified for location analysis.

[0055] As we know from the principle of the AAC encoder, AAC audio is encoded frame by frame. For each frame of raw audio data, the encoder converts it into frequency domain information through a filter bank and MDCT transform. This frequency domain information is then quantized to obtain 1024 QMDCT coefficients. In other words, during the time-frequency conversion process, the AAC encoder divides the frequency range of the audio from 0 to the sampling rate value into 1024 frequency bands, and each QMDCT coefficient represents the quantized information of that frame of audio in the corresponding frequency band. Therefore, if we can calculate the frequency domain information of the AAC audio and find the masked portion from the 1024 frequency bands, we can modify the QMDCT coefficients corresponding to that frequency band to embed secret information. In this case, since the frequency band corresponding to the modified QMDCT coefficient is masked, meaning the modification is to the masked signal in the audio, it is imperceptible to the human ear. This, to a certain extent, ensures the auditory stealth of the steganography algorithm.

[0056] To find the masked QMDCT coefficients in each frame of audio, we can... Figure 4 The process involves calculation.

[0057] Depend on Figure 4 As can be seen, when calculating the masked frequency band, a Short-Time Fourier Transform (STFT) is first required on the original audio. The spectrum obtained after the audio undergoes the STFT is shown below. Figure 5 As shown.

[0058] Next, the STFT coefficients of each frame of audio need to be calculated and converted into sound pressure level (SPL) to better represent the signal strength of that frequency band. During the conversion, the frequency band index of the peak SPL is recorded. Then, the masking coefficient of each peak signal on all other signals is calculated. It's important to note that since the masking effect mainly occurs within a narrow band of the peak signal's frequency, the masking coefficient of the peak signal in that frequency band decreases rapidly as the distance between two frequency bands increases. Finally, the masking coefficients of all peak signals in each frequency band are summed to obtain the masking threshold for that frequency band. Finally, the masking threshold and SPL of each frequency band are compared. If the signal strength of that frequency band is lower than the masking threshold, it means that the sound intensity of that signal is less than the masking effect value of other peak signals on that frequency band, and thus that frequency band can be considered a masked frequency band. After obtaining the masked frequency bands of the original audio, we also know which positions of the QMDCT coefficient values ​​can be modified. Under different preset thresholds H, the modifiable positions of the QMDCT coefficients in the original audio are as follows: Figure 6 As shown.

[0059] exist Figure 6 In the diagram, the horizontal axis represents the number of frames in the carrier audio, and the vertical axis represents the frequency band of each frame. Yellow dots represent frequencies in that frame that are not masked and therefore cannot be modified, while blue dots represent positions that can be modified. It can be seen that when the threshold H is 16, 10, and 5, the number of modifiable positions in the original carrier audio is not large, and they are mainly concentrated in the high-frequency domain. However, when the threshold H is 0, the number of modifiable positions in the original carrier audio increases geometrically, reaching almost 90% of the original carrier audio. This is because when the threshold H is 0, it is equivalent to retaining only the positions where the peak signal is located.

[0060] 1.2 Design of the distortion cost function.

[0061] First, statistical analysis was performed on the QMDCT coefficients of the original audio. The frequency and frequency histogram of its first-order statistical distribution are shown below. Figure 7 As shown.

[0062] As can be seen, although the QMDCT coefficients range from [-12, 11], most are concentrated in the interval [-2, 2], with over 70% of the QMDCT coefficients having a value of 0, and over 90% having a value in the range [-1, 1]. This means that without further restrictions on the modification location, modifications will likely be concentrated in the range [-1, 1], thus altering the statistical characteristics of the QMDCT coefficients. To address this, we conducted the following small experiment. First, we calculated the masked frequency bands in each audio frame using the method mentioned in the previous section. Then, we designed a distortion cost function as shown in Formula 1-2 for each frequency band. Next, we embedded secret information using the STC framework. Finally, we performed statistical analysis on the modified QMDCT coefficients to observe their corresponding statistical characteristics.

[0063]

[0064] Where cost represents the single-point distortion at this location, ∞ represents infinity, and i represents the frequency band index of the QMDCT coefficient, with a value range of {0, 1, ..., 1023}. mod[i] = 0 indicates that the frequency band is not a masked frequency band and is an unmodifiable location, while mod[i] = 1 indicates that the frequency band is a masked frequency band and can be modified. Therefore, the distortion at this location is set to a very small constant of 0.01. After steganography embedding under the guidance of this distortion function, the distribution characteristics of the QMDCT coefficient values ​​before and after modification are compared, and the results are as follows: Figure 8 As shown.

[0065] As can be seen from the figure, the STC steganography framework mainly modifies the QMDCT coefficients in the interval [-1, 1]. However, due to the addition and subtraction of 1, the modified QMDCT coefficient values ​​are still in the interval [-1, 1]. Therefore, overall, even after the modified embedding, the number of QMDCT coefficients with values ​​of [-1, 1] is still more than 90%, which maintains the original statistical distribution characteristics to a certain extent, demonstrating the feasibility of the distortion cost function.

[0066] II. The following section introduces the steganography algorithm that uses the above principles.

[0067] This invention analyzes the masking effect in AAC encoding, divides QMDCT coefficients into masked and unmasked parts, designs a corresponding distortion cost function, modifies the evolved QMDCT coefficients using the STC steganography framework, embeds secret information, and proposes an AAC secure steganography algorithm based on the masking effect.

[0068] First, the masking thresholds for each frequency band during AAC encoding are calculated and analyzed. Based on these thresholds, QMDCT coefficients are classified, and those below the threshold are recorded as modifiable QMDCT coefficients. Then, combining the distortion function and the STC framework, secret information is embedded. The steganography algorithm framework is as follows: Figure 9 As shown.

[0069] The specific embedding process is as follows:

[0070] S1. Input the original WAV audio into the AAC encoder and encode it according to the normal procedure. Obtain the complete QMDCT coefficient sequence Q: {Q1, Q2, ..., Q} in the quantization module. n};

[0071] S2. While acquiring the QMDCT coefficient sequence, process the original WAV audio to obtain the masked frequency bands in each frame of audio and record them as modifiable positions.

[0072] S3. Calculate the single-point distortion cost of the QMDCT coefficients for each frequency band according to the pre-designed distortion cost function, and obtain the distortion cost sequence Cost: {Cost1, Cost2, ..., Cost...} n};

[0073]

[0074] S4. Convert the secret message m into a binary bit stream sequence and scramble it using the key key to obtain m′; at the same time, convert the QMDCT coefficient sequence Q obtained in step 1 into a binary sequence Q′.

[0075] S5. Input the secret message sequence m′, QMDCT coefficient sequence Q′ and distortion cost sequence Cost into the STC framework to obtain the embedding method with the minimum overall distortion, and output the final secret-containing QMDCT carrier sequence P′.

[0076] S6. The carrier sequence P′ is converted into a QMDCT coefficient sequence P through an inverse mapping operation, and the QMDCT coefficients are replaced in the quantization module of the second encoding process. Then, noiseless encoding and other processes are performed to finally obtain the encrypted AAC audio and complete the steganography of secret information.

[0077] III. Steganography Experiment.

[0078] Step 1: Input PCM samples and use the steganography method proposed in this invention to embed random secret information during the AAC speech compression coding process to generate Stego-encrypted speech.

[0079] Step 2: Steganography of experimental results.

[0080] To verify the performance of the algorithm of this invention, this invention tests the steganography algorithm from three aspects: auditory concealment, statistical security, and embedding capacity.

[0081] (1) Auditory concealment experiment.

[0082] In the auditory concealment experiment, the steganography algorithm designed in this invention was compared with two other AAC steganography algorithms based on the QMDCT coefficient domain. PEAQ was used to evaluate the sound quality of the condensed audio. The experimental results are shown below.

[0083] Table 1.1 ODG values ​​of three steganography algorithms under different payloads

[0084]

[0085] As shown in Table 1.1, based on the ODG parameters, the imperceptibility of the three steganography methods, from highest to lowest, is: the algorithm presented in this paper > the steganography algorithm. [2] Steganography algorithm [1] This is because of the steganography algorithms and steganography algorithms discussed in this chapter. [2] All the steganography algorithms presented are adaptive steganography algorithms. Before embedding, they select the embedding scheme with the minimum overall distortion from the audio through a distortion cost function and the STC framework. Therefore, the auditory concealment experiment shows that the steganography algorithm designed in this invention has little impact on the quality of dense speech, and the steganized speech still has high speech quality. Thus, the steganography algorithm proposed in this invention has good auditory concealment.

[0086] (2) Hidden capacity experiment.

[0087] In the steganography capacity experiment, 4000 original audio files were embedded into each of the three steganography algorithms. The average embedding capacity was then calculated, and the final experimental results are shown in Table 1.2.

[0088] Table 1.2 Embedding capacity of three steganography algorithms

[0089]

[0090] Through capacity statistics experiments, the hidden capacity of the steganography algorithm in this paper was analyzed, verifying that the algorithm has a high hidden capacity.

[0091] (3) Statistical safety experiment.

[0092] In the statistical security experiment, steganalysis experiments will be conducted on three audio steganography algorithms. Two steganalysis algorithms will be used for testing in the experiment. The security of the corresponding steganography algorithm will be evaluated by the discrimination accuracy of the steganalysis algorithm, as shown in Equation 1.

[0093]

[0094] In this table, ACC represents the discrimination accuracy of the steganalysis algorithm, TNR refers to the true negative rate, which indicates whether the steganalysis algorithm correctly predicted whether the audio was cryptic, and TPR is the true positive rate, which indicates whether the steganalysis algorithm correctly predicted whether the audio was the original audio. The final experimental results are shown in Table 1.3.

[0095] Table 1.3 Statistical security experimental results of three steganography algorithms

[0096]

[0097] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. An AAC steganography algorithm based on masking effect, characterized in that, comprising The original audio is encoded to obtain a QMDCT coefficient sequence , and converted to obtain a binary sequence ; The original audio is transformed to obtain a masked frequency segment in each frame of audio, and the masked frequency segment is taken as a modifiable position; a single-point distortion cost of a QMDCT coefficient of each frequency segment is calculated based on the obtained modifiable position by using a distortion cost function, to obtain a distortion cost sequence ; the distortion cost sequence is calculated using the following equation ； wherein mod[i] is 0 means that the frequency band is not a masked frequency band, mod[i] is 1 means that the frequency band is a masked frequency band; The secret message sequence , the QMDCT coefficient sequence and the distortion cost sequence In the input STC framework, the embedding method with the minimum overall distortion is obtained, and the final QMDCT carrier sequence containing the secret is output .

2. The AAC steganography algorithm based on masking effect according to claim 1, characterized in that, The original wav audio input is encoded into the AAC encoder, transformed by MDCT and the complete sequence of QMDCT coefficients is obtained in the quantization module .

3. The AAC secure steganography algorithm based on masking effect according to claim 2, characterized in that, the spectrum diagram obtained by performing short-time Fourier transform on the original audio, calculating the STFT coefficients of each frame of audio, converting them into sound pressure levels, and calculating the masking coefficients of each peak signal to all other signals; calculating the masking thresholds of all frequency bands through the masking coefficients; comparing the masking thresholds and the sound pressure levels of each frequency band, and if the signal intensity of the frequency band is lower than the masking threshold, the frequency band is a masked frequency band.

4. The AAC steganography algorithm based on masking effect according to claim 3, characterized in that, In each frequency band, the masking threshold is the sum of the masking coefficients of all peak signals.

5. The AAC steganography algorithm based on masking effect according to claim 4, characterized in that, The secret message m is converted into a binary bit stream sequence, and is subjected to disorder processing with the key key, to obtain .

6. The AAC secure steganography algorithm based on masking effect according to claim 5, characterized in that, The vector sequence is obtained through inverse mapping operations. Convert to QMDCT coefficient sequence P; performing QMDCT coefficient replacement in the quantization module of the second encoding process; encoding the AAC audio containing the secret according to the replaced QMDCT coefficients.

7. A system for masking effect based AAC steganography, characterized in that, comprising The first module is configured to encode the original audio to obtain a QMDCT coefficient sequence , and convert the QMDCT coefficient sequence to obtain a binary sequence ; The second module is configured to transform the original audio to obtain a masked frequency segment in each frame of audio, and take the masked frequency segment as a modifiable position; calculate a single-point distortion cost of a QMDCT coefficient of each frequency segment based on the obtained modifiable position by using a distortion cost function, and obtain a distortion cost sequence ; the distortion cost sequence is calculated using the following equation ； wherein mod[i] is 0 means that the frequency band is not a masked frequency band, mod[i] is 1 means that the frequency band is a masked frequency band; The third module: configured to transmit secret message sequences QMDCT coefficient sequence and distortion cost sequence Inputting the data into the STC framework yields the embedding method with minimal overall distortion, and outputs the final QMDCT vector sequence containing density. .

Images