Abstract:The use of photo-response non-uniformity (PRNU) has been proposed as the basis of a sensor fingerprint for common source camera identification. We perform tests of the PRNU-based fingerprint on a set of videos chosen to represent a wide range of potential inputs. Based on the results of these tests, we propose a confidence weighting scheme to address the problem of extracting a viable fingerprint from videos where high-frequency content (e.g. edges) persist at a given image location. We further show that the e… Show more
“…The results show that among the 13 schemes evaluated, the top scheme is G4-3 which combines the confidence weighting method proposed by McCloskey [13] and the ZM and WF operations proposed by Chen et al [3], achieving a TPR of 84.79%. It is also clear from Table 4 that some other schemes (G2-4,G3-2,G3-3,G3-4,G4-1,G4-2,G4-4) also perform well, achieving TPRs between 82.27% and 84.79%.…”
Section: Which Is the Most Effective?mentioning
confidence: 95%
“…Since some critical information for implementation is implicit in [2], we are not including it in our evaluation but use [13] to represent this line of work. The weight w for each pixel p is calculated in [13] as:…”
Section: Enhancing the Spn Of A Test Imagementioning
confidence: 99%
“…This scheme weights against those image regions that we have less confidence in its SPN, typically being edge and highly textured regions. For example, image gradient magnitudes are exploited in [13] while a pair of intensity and texture features are used in [2]. Since some critical information for implementation is implicit in [2], we are not including it in our evaluation but use [13] to represent this line of work.…”
Section: Enhancing the Spn Of A Test Imagementioning
confidence: 99%
“…For example, image gradient magnitudes are exploited in [13] while a pair of intensity and texture features are used in [2]. Since some critical information for implementation is implicit in [2], we are not including it in our evaluation but use [13] to represent this line of work. The weight w for each pixel p is calculated in [13] as:…”
Section: Enhancing the Spn Of A Test Imagementioning
confidence: 99%
“…For completeness, we include both implementations in our experiments. Confidence weighting [13]: While the attenuation model in [12] decides the weight of a component simply by its magnitude, there are a few works that propose to use the content adaptive weighting scheme. This scheme weights against those image regions that we have less confidence in its SPN, typically being edge and highly textured regions.…”
Section: Enhancing the Spn Of A Test Imagementioning
The sensor pattern noise (SPN) based source camera identification technique has been well established. The common practice is to subtract a denoised image from the original one to get an estimate of the SPN. Various techniques to improve SPN's reliability have previously been proposed. Identifying the most effective technique is important, for both researchers and forensic investigators in law enforcement agencies. Unfortunately, the results from previous studies have proven to be irreproducible and incomparable -there is no consensus on which technique works the best. Here, we extensively evaluate various ways of enhancing the SPN by using the public "Dresden" database. We identify which enhancing methods are more effective and offer some insights into the behavior of SPN. For example, we find that the most effective enhancing methods share a common strategy of spectrum flattening. We also show that methods that only aim at reducing the contamination from image content do not lead to satisfying results, since the non-unique artifacts (NUA) among different cameras are the major troublemaker to the identification performance. While there is a trend of employing sophisticate methods to predict the impact of image content, our results suggest that more effort should be invested to tame the NUAs.
“…The results show that among the 13 schemes evaluated, the top scheme is G4-3 which combines the confidence weighting method proposed by McCloskey [13] and the ZM and WF operations proposed by Chen et al [3], achieving a TPR of 84.79%. It is also clear from Table 4 that some other schemes (G2-4,G3-2,G3-3,G3-4,G4-1,G4-2,G4-4) also perform well, achieving TPRs between 82.27% and 84.79%.…”
Section: Which Is the Most Effective?mentioning
confidence: 95%
“…Since some critical information for implementation is implicit in [2], we are not including it in our evaluation but use [13] to represent this line of work. The weight w for each pixel p is calculated in [13] as:…”
Section: Enhancing the Spn Of A Test Imagementioning
confidence: 99%
“…This scheme weights against those image regions that we have less confidence in its SPN, typically being edge and highly textured regions. For example, image gradient magnitudes are exploited in [13] while a pair of intensity and texture features are used in [2]. Since some critical information for implementation is implicit in [2], we are not including it in our evaluation but use [13] to represent this line of work.…”
Section: Enhancing the Spn Of A Test Imagementioning
confidence: 99%
“…For example, image gradient magnitudes are exploited in [13] while a pair of intensity and texture features are used in [2]. Since some critical information for implementation is implicit in [2], we are not including it in our evaluation but use [13] to represent this line of work. The weight w for each pixel p is calculated in [13] as:…”
Section: Enhancing the Spn Of A Test Imagementioning
confidence: 99%
“…For completeness, we include both implementations in our experiments. Confidence weighting [13]: While the attenuation model in [12] decides the weight of a component simply by its magnitude, there are a few works that propose to use the content adaptive weighting scheme. This scheme weights against those image regions that we have less confidence in its SPN, typically being edge and highly textured regions.…”
Section: Enhancing the Spn Of A Test Imagementioning
The sensor pattern noise (SPN) based source camera identification technique has been well established. The common practice is to subtract a denoised image from the original one to get an estimate of the SPN. Various techniques to improve SPN's reliability have previously been proposed. Identifying the most effective technique is important, for both researchers and forensic investigators in law enforcement agencies. Unfortunately, the results from previous studies have proven to be irreproducible and incomparable -there is no consensus on which technique works the best. Here, we extensively evaluate various ways of enhancing the SPN by using the public "Dresden" database. We identify which enhancing methods are more effective and offer some insights into the behavior of SPN. For example, we find that the most effective enhancing methods share a common strategy of spectrum flattening. We also show that methods that only aim at reducing the contamination from image content do not lead to satisfying results, since the non-unique artifacts (NUA) among different cameras are the major troublemaker to the identification performance. While there is a trend of employing sophisticate methods to predict the impact of image content, our results suggest that more effort should be invested to tame the NUAs.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.