Academia.eduAcademia.edu

Figure 4 – uploaded by Marco Fontani

See full PDFdownloadDownload figure
BRAND CLASSIFICATION CONFUSION MATRIX ON UNKNOWN DEVICES TABLE 5

Table 5 BRAND CLASSIFICATION CONFUSION MATRIX ON UNKNOWN DEVICES TABLE 5

Related Figures (5)

Fig. 1. Comparison between fragments of file container of an original video acquired by a Samsung Galaxy S3 (a) and of the same video after its editing through FFmpeg (b).
Fig. 1. Comparison between fragments of file container of an original video acquired by a Samsung Galaxy S3 (a) and of the same video after its editing through FFmpeg (b).
Fig. 2. Summary of statistics for the device D03 Huawei P9: Doo, DEX FTool DF! RS pyouPube. DWhatsApp and Doq separately restricted to Apple Devices and to other Android Devices. Summ: of statistics for the device D15 Apple iPhone 6: : Doo, Fig. 3. if Tool F WhatsA DEF 2° Dog 4, pyourube’ Doa *??, and Doa separately restricted to other Apple Devices and to Android Devices.
Fig. 2. Summary of statistics for the device D03 Huawei P9: Doo, DEX FTool DF! RS pyouPube. DWhatsApp and Doq separately restricted to Apple Devices and to other Android Devices. Summ: of statistics for the device D15 Apple iPhone 6: : Doo, Fig. 3. if Tool F WhatsA DEF 2° Dog 4, pyourube’ Doa *??, and Doa separately restricted to other Apple Devices and to Android Devices.
VIDEO INTEGRITY VERIFICATION SUMMARY: DEVICE ID (JD); MIN AND MAX VALUES OF INTRA-CLASS VARIABILITY Doo (Original); MIN AND MAX VALUES OF INTER-CLASS VARIABILITY Doa (Altered); AUC VALUES SUMMARISING THE DISTINGUISHABILITY BETWEEN Doo AND Doa (AUC-AIt); MII AND MAX VALUES OF THE INTER-CLASS STATISTICS Dog (Other Devs); AUC VALUES SUMMARISING RESPECTIVELY THE DISTINGUISHABILITY BETWEEN THE CLASSES Doo AND Dog (AUC-Devs), AND THE CLASSES Doo AND Do 4 = Dog U Dog (AUC-AII).
VIDEO INTEGRITY VERIFICATION SUMMARY: DEVICE ID (JD); MIN AND MAX VALUES OF INTRA-CLASS VARIABILITY Doo (Original); MIN AND MAX VALUES OF INTER-CLASS VARIABILITY Doa (Altered); AUC VALUES SUMMARISING THE DISTINGUISHABILITY BETWEEN Doo AND Doa (AUC-AIt); MII AND MAX VALUES OF THE INTER-CLASS STATISTICS Dog (Other Devs); AUC VALUES SUMMARISING RESPECTIVELY THE DISTINGUISHABILITY BETWEEN THE CLASSES Doo AND Dog (AUC-Devs), AND THE CLASSES Doo AND Do 4 = Dog U Dog (AUC-AII).
BRAND CLASSIFICATION CONFUSION MATRIX, AND PERCENTAGE OF CORRECT CLASSIFICATION IN THE TOP 3 As to the computational effort, the average time required for a brand classification of a video using the proposed technique is just 7.74 seconds on a dataset with eight different brands.
BRAND CLASSIFICATION CONFUSION MATRIX, AND PERCENTAGE OF CORRECT CLASSIFICATION IN THE TOP 3 As to the computational effort, the average time required for a brand classification of a video using the proposed technique is just 7.74 seconds on a dataset with eight different brands.
VIAIN FEATURES OF THE VISION DATASET. MORE DETAILS IN [27]. different devices, an iPad Air mini (iOS 11.0.1), an iPhone 7 GOS 10.3.1) and an iPhone X (GOS 11.2.1), that are not included in the VISION dataset. For each camera, we captured five videos and we tried to assess the capturing device brand (as defined in the previous experiment), by exploiting only the VISION dataset as the training set. As Table 5 shows, the brand for the iPad Air mini and for the iPhone 7 are correctly determined. Conversely, iPhone X videos are not assigned to any known brand. This is not surprising: while the first two models are expected to have some similarities with other Apple containers, the iPhone X features the new HEVC codec which revolutionizes the container structure. This experiment highlights that, in certain cases, the proposed tool can be exploited even if we do not have references from the same model in the training set.
VIAIN FEATURES OF THE VISION DATASET. MORE DETAILS IN [27]. different devices, an iPad Air mini (iOS 11.0.1), an iPhone 7 GOS 10.3.1) and an iPhone X (GOS 11.2.1), that are not included in the VISION dataset. For each camera, we captured five videos and we tried to assess the capturing device brand (as defined in the previous experiment), by exploiting only the VISION dataset as the training set. As Table 5 shows, the brand for the iPad Air mini and for the iPhone 7 are correctly determined. Conversely, iPhone X videos are not assigned to any known brand. This is not surprising: while the first two models are expected to have some similarities with other Apple containers, the iPhone X features the new HEVC codec which revolutionizes the container structure. This experiment highlights that, in certain cases, the proposed tool can be exploited even if we do not have references from the same model in the training set.
Related topics:
EngineeringComputer Science
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved