Research On ENF Signal-based Blind Detection Of Digital Audio Forgery

While digital audio recording has become much easier now with various digitalmultimedia devices, the forgery of digital audio has become more accessible due tothe variety of editing softwares. It therefore necessitates reliable approaches to verifythe originality, integrity and authenticity of digital audio. The forgery detectionmethods based on electric network frequency (ENF), though highlighted in recentyears, exhibit critical limitations. For example, existing ENF based methods requirean ENF reference database for detection, which is not easily accessible in mostapplications. Besides, most existing methods can not locate the forgery precisely. Tosolve these problems, we investigate blind audio forgery detection methods withoutENF reference database. The major contributions of this thesis are as follows:(1) An approach based on the maximum offset for cross correlation (MOCC) ofENF signal is proposed for audio forgery detection. We divide the ENF signal intoblocks and consider the first block as the reference signal. We introduce the MOCC asthe offset at which the cross-correlation between the reference signal and each blockreaches its maximum. According to the consistency of the MOCCs, it is possible totell whether an audio clip has undergone forgery, what type of forgery it is as well aswhere the forgery is located. For different applications, we propose three methodswhich are visual detection, automatic detection and fast detection respectively.Experiments show that the proposed MOCC-based method accurately estimates theforgery location and the forgery type. It is worth mentioning that since the proposedmethod operates in time domain rather than transform domain, the computationalcomplexity is lower than existing methods based on the phase information of ENFsignals.(2) An improved method is proposed for the method in (1) which is subject tonoise interference in the calculation of MOCC. An ideal sine wave is introduced asthe reference signal. The ENF signal is enhanced exploiting the multiplecross-correlations between the reference signal and the sub-blocks. Then the MOCC is calculated for each sub-block of the enhanced ENF signal with the ideal reference.According to the inconsistency of the MOCCs we can locate the forgery area anddetermine the forgery type. Experiments show that the proposed method caneffectively increase the SNR of the ENF signal, increasing reliability of the obtainedMOCCs. Higher detection accuracy can be achieved under low false alarm rate. Inaddition, the method shows robustness in terms of resistance to the additive noise,audio re-sampling and audio compression.(3) In order to achieve higher precision when locating the forgery region, adual judgment mechanism based on the minimum offset for average magnitudedifference (MOAMD) is proposed. MOAMD is employed for audio forgery detection,which simplifies the calculation of MOCC. Then the forgery region is located with thedual mechanism composed of investigating the changes of the MOAMD curve as wellas checking the slope of the extreme points of the MOAMD curve. Experiments showthat the precision of the forgery region locating is increased with the proposedmethod. In addition, the second part of the dual mechanism can also be extended toother methods.(4) An audio forgery detection approach is proposed based on the ENFcross-correlation coefficients in neighborhood and the fast transversal filter (FTF).The cross-correlation coefficients of neighboring ENF sub-blocks are calculated. Theadaptive FTF is then applied to the obtained coefficients. According to the changes ofthe error energy after filtering, we can detect forgery and locate the forgery region.Experiments establish that the proposed method advantages the forgery detectionespecially when the ENF fluctuates in a large range and has low SNR.