Limits...
Existence detection and embedding rate estimation of blended speech in covert speech communications.

Li L, Gao Y - Springerplus (2016)

Bottom Line: The average zero crossing rate (ZCR) is calculated for each OED frame, and the minimum average ZCR and AZCR-OED of the entire speech signal are extracted as features.The results demonstrate that without attack, the detection accuracy can reach 80 % or more when the embedding rate is greater than 10 %, and the estimated embedding rate is similar to the real value.And when some attacks occur, it can also reach relatively high detection accuracy.

View Article: PubMed Central - PubMed

Affiliation: College of Electronics and Information Engineering, Sichuan University, Chengdu, 610064 Sichuan China.

ABSTRACT
Covert speech communications may be used by terrorists to commit crimes through Internet. Steganalysis aims to detect secret information in covert communications to prevent crimes. Herein, based on the average zero crossing rate of the odd-even difference (AZCR-OED), a steganalysis algorithm for blended speech is proposed; it can detect the existence and estimate the embedding rate of blended speech. First, the odd-even difference (OED) of the speech signal is calculated and divided into frames. The average zero crossing rate (ZCR) is calculated for each OED frame, and the minimum average ZCR and AZCR-OED of the entire speech signal are extracted as features. Then, a support vector machine classifier is used to determine whether the speech signal is blended. Finally, a voice activity detection algorithm is applied to determine the hidden location of the secret speech and estimate the embedding rate. The results demonstrate that without attack, the detection accuracy can reach 80 % or more when the embedding rate is greater than 10 %, and the estimated embedding rate is similar to the real value. And when some attacks occur, it can also reach relatively high detection accuracy. The algorithm has high performance in terms of accuracy, effectiveness and robustness.

No MeSH data available.


Statistical results for  and , the AZCR-OED of blended speech when the odd–even points are aligned and inverted, respectively
© Copyright Policy - OpenAccess
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4940342&req=5

Fig3: Statistical results for and , the AZCR-OED of blended speech when the odd–even points are aligned and inverted, respectively

Mentions: Experiment 2: we made five copies of the carrier speech group. Then, secret speech was embedded into the carrier speech signal using the blending-based speech hiding algorithm with five different embedding rates. Because the embedding rate is typically high in practical applications, we selected 10, 30, 50, 70, and 100 % as the embedding rates in the experiment. When the odd–even points are aligned, the AZCR-OED of blended speech is unrelated to the hidden degree factor. Thus, we used a hidden degree factor of 0.1 in the experiment. Consequently, we obtained five blended speech groups with different embedding rates. We calculated the OED of each blended speech signal and the corresponding average ZCR . Then, we inverted the odd–even points of each blended speech, and calculated the OED of each inverted blended speech signal as well as the corresponding average ZCR . Figure 3 shows the statistical results.Fig. 3


Existence detection and embedding rate estimation of blended speech in covert speech communications.

Li L, Gao Y - Springerplus (2016)

Statistical results for  and , the AZCR-OED of blended speech when the odd–even points are aligned and inverted, respectively
© Copyright Policy - OpenAccess
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4940342&req=5

Fig3: Statistical results for and , the AZCR-OED of blended speech when the odd–even points are aligned and inverted, respectively
Mentions: Experiment 2: we made five copies of the carrier speech group. Then, secret speech was embedded into the carrier speech signal using the blending-based speech hiding algorithm with five different embedding rates. Because the embedding rate is typically high in practical applications, we selected 10, 30, 50, 70, and 100 % as the embedding rates in the experiment. When the odd–even points are aligned, the AZCR-OED of blended speech is unrelated to the hidden degree factor. Thus, we used a hidden degree factor of 0.1 in the experiment. Consequently, we obtained five blended speech groups with different embedding rates. We calculated the OED of each blended speech signal and the corresponding average ZCR . Then, we inverted the odd–even points of each blended speech, and calculated the OED of each inverted blended speech signal as well as the corresponding average ZCR . Figure 3 shows the statistical results.Fig. 3

Bottom Line: The average zero crossing rate (ZCR) is calculated for each OED frame, and the minimum average ZCR and AZCR-OED of the entire speech signal are extracted as features.The results demonstrate that without attack, the detection accuracy can reach 80 % or more when the embedding rate is greater than 10 %, and the estimated embedding rate is similar to the real value.And when some attacks occur, it can also reach relatively high detection accuracy.

View Article: PubMed Central - PubMed

Affiliation: College of Electronics and Information Engineering, Sichuan University, Chengdu, 610064 Sichuan China.

ABSTRACT
Covert speech communications may be used by terrorists to commit crimes through Internet. Steganalysis aims to detect secret information in covert communications to prevent crimes. Herein, based on the average zero crossing rate of the odd-even difference (AZCR-OED), a steganalysis algorithm for blended speech is proposed; it can detect the existence and estimate the embedding rate of blended speech. First, the odd-even difference (OED) of the speech signal is calculated and divided into frames. The average zero crossing rate (ZCR) is calculated for each OED frame, and the minimum average ZCR and AZCR-OED of the entire speech signal are extracted as features. Then, a support vector machine classifier is used to determine whether the speech signal is blended. Finally, a voice activity detection algorithm is applied to determine the hidden location of the secret speech and estimate the embedding rate. The results demonstrate that without attack, the detection accuracy can reach 80 % or more when the embedding rate is greater than 10 %, and the estimated embedding rate is similar to the real value. And when some attacks occur, it can also reach relatively high detection accuracy. The algorithm has high performance in terms of accuracy, effectiveness and robustness.

No MeSH data available.